MIA Literature Reading - Latest Progress in Medical Image Processing in Chronic Kidney Disease Assessment [2021]

0 abstract

Accurate assessment of kidney function and structure is crucial for the diagnosis and prognosis of chronic kidney disease (CKD). Advanced imaging techniques, including magnetic resonance imaging (MRI), ultrasound elastography (UE), computed tomography (CT) and scintigraphy (PET, SPECT), provide the opportunity to non-invasively retrieve structural, functional and molecular information that can detect Changes in renal tissue properties and function. Artificial intelligence's ability to transform traditional medical imaging into fully automated diagnostic tools is currently being extensively studied. In addition to qualitative analysis of renal medical imaging, texture analysis combined with machine learning techniques as a quantification of renal tissue heterogeneity provides a promising complementary tool for renal function decline prediction. Interestingly, deep learning has the potential to become a new approach to kidney function diagnosis. This article presents a survey covering qualitative and quantitative analyzes applied to new medical imaging technologies for monitoring renal function decline. First, we summarize the use of different medical imaging modalities to monitor CKD . Then, we demonstrate the ability of artificial intelligence (AI) to guide renal function assessment, from segmentation to disease prediction, discussing how texture analysis and machine learning techniques have been used in recent clinical trials. Emerging from research to improve kidney function monitoring and prediction . This article reviews the role of artificial intelligence in kidney segmentation .

1 Introduction

Chronic kidney disease (CKD) is one of the major public health challenges considered by the World Health Organization (WHO) today. In France, it affects nearly 82,000 people, with the number of affected patients increasing by 2% every year (Bayat et al, 2010). CKD remains a widespread and important public health problem, affecting more than 12% of the global population (Jiang and Lerman, 2019). It is characterized by progressive and irreversible deterioration of renal function, accompanied by low glomerular filtration rate (GFR), leading to end-stage renal disease, in which the kidneys are completely damaged and unable to filter blood, requiring renal dialysis or Kidney transplantation (Jiang and Lerman, 2019).

Therefore, when the disease reaches the end stage (end-stage chronic renal insufficiency - IRCT), two options are possible . The most common is hemodialysis . This blood filtration technology still has limitations. Patients must undergo dialysis three times a week for four hours each time. Another solution available for CRTI patients is kidney transplantation . It is generally accepted that transplantation provides a better quality of life than dialysis. In fact, a transplant makes it possible to find an almost normal kidney activity while maintaining a nearly "normal" lifestyle. Early detection of CKD provides the ability to guide patient management and reduce mortality by preventing progression to endpoints, which are often associated with numerous health complications, including heart and bone disease and hypertension (Saha et al., 2019).

The kidneys filter minerals and metabolites from the blood as urine (notohamiprojo et al., 2010). CKD is accompanied by the accumulation of extracellular matrix (ECM) or fibrosis-related deposits that alter the way tissues handle water, as well as tissue stiffness and macromolecule content (Notohamiprojo et al., 2010). Loss of peritubular capillaries results in reduced renal oxygenation and perfusion (Nangaku, 2006). Renal function is typically assessed by estimating glomerular filtration rate (eGFR) based on serum creatine levels (sCr) or gold standard biopsy followed by histopathological analysis . However, these medical practices have some flaws and limitations . sCr is a late indicator of renal function decline and requires blood sample analysis. Furthermore, the GFR estimating equation lacks precision and cannot provide split kidney function measurements (Notohamiprodjo et al., 2010). Although biopsy is the gold standard method for assessing kidney microstructure and quantifying the cause of renal dysfunction, it has other shortcomings. First, the invasive procedure is associated with bleeding and possible pain for the patient, thus limiting follow-up evaluation. Second, biopsy is subject to sampling bias as it samples less than 1% of a kidney, and the medulla is typically not included as it is limited in spatial resolution (Baues et al., 2017; Leung et al. , 2017).

Radiology can provide structural and functional markers that have been reported to be helpful in predicting and following up renal insufficiency. It plays an increasingly important role in the assessment of renal function . Table 1 summarizes recently used medical imaging modalities . Magnetic resonance imaging (MRI) has the potential to assess renal microstructural organization, diffusion, perfusion, and oxygenation. Additionally, it provides the ability to measure renal hemodynamics, quantify tissue relaxation time, macromolecules and elasticity, and characterize renal tissue metabolites.

Ultrasound elastography (UE) allows the identification of structural information regarding the mechanical properties of tissues. Computed tomography (CT) can provide anatomical and functional information, but it is limited to X-ray exposure and contrast injection. Scintigraphy combines functional and perfusion measurements using radioactive tracers. Qualitative analysis of these medical images shows a promising tool for assessing kidney function .

Interestingly, the quantification of medical images by analyzing image pixels , also known as texture analysis, has been studied as a practical complementary tool. Texture analysis is derived from mathematical equations, studying the spatial arrangement of grayscale pixels and revealing relationships between them that are often invisible to the human eye. Considering the impact of renal structure and renal disease on functional marker distribution, texture has the potential to reflect histopathological heterogeneity (Ding et al, 2019; Shi et al, 2018). Combining texture analysis with traditional machine learning methods expands medical imaging capabilities by providing additional numerical descriptors that can be used to diagnose and predict renal dysfunction.

In medical image processing, quantization refers to mapping the pixel values ​​of an image to a discrete set of values. Typically, medical images are presented in the form of grayscale images, with each pixel representing the grayscale value of a specific area. The quantization process converts continuous pixel grayscale values ​​into discrete numerical values, which facilitates subsequent image analysis and processing. The purpose of quantization is to reduce the representation complexity of image data so that images can be better stored, transmitted and processed. Common quantization methods include simple linear quantization and nonlinear quantization. Linear quantization divides the pixel grayscale range into several intervals evenly, and then maps each interval to a specific discrete value. Nonlinear quantization performs mapping based on the distribution characteristics of pixel gray values, such as using histogram equalization and other methods.

Most importantly, deep learning has recently been investigated in clinical applications aimed at diagnosing kidney function and chronic kidney disease patients. It is a branch of machine learning inspired by biological neurons, which are composed of multiple layers of interconnected nodes, providing a new powerful tool for image feature extraction and patient classification.

In this survey, we first review the application of medical imaging techniques in the assessment of renal dysfunction, specifically CKD. We then demonstrate the potential of AI in improving renal function prediction and diagnostic performance by summarizing recent applications of texture and machine learning techniques, including neural networks, on different medical imaging modalities. The role of artificial intelligence in kidney segmentation is also discussed, as accurate identification of kidneys on medical images is an important step to save time and transform renal parenchyma delineation into a subject-independent problem as it requires input from human experts or computer-aided diagnosis carried out before.

2 Magnetic resonance imaging (MRI)

2.1 Diffusion magnetic resonance imaging

2.1.1 Diffusion-weighted imaging (DWI)

Renal fibrosis, characterized by extracellular matrix (ECM) deposition, plays a key role in the development of CKD (Leung et al., 2017; Nangaku, 2006). The mobility of water molecules within tissues is thought to be reduced due to the involvement of tissue cells. Therefore, water diffusion headers can reflect tissue microstructure. Diffusion-weighted imaging (DWI) is a magnetic resonance modality that uses the motion of water molecules as contrast to provide in vivo measurements of water diffusion, or Brownian motion. It uses powerful bipolar magnetic gradients to create sensitivity of the received signal to water movement, thereby describing the way tissues confine water (Notohamiprodjo et al., 2010). Apparent diffusion coefficient (ADC) is a DWI biomarker that corresponds to a global measure of water diffusion and microcirculation within tissues (notohamiprojo et al, 2010). Le Bihan et al., 1988) introduced the intravoxel incoherent motion (IVIM) model in order to distinguish true diffusion or molecular water diffusion due to blood flow in capillary networks from pseudodiffusion or perfusion. Parameters derived from IVIM are true diffusion (D) related to extravascular water molecule movement or pure diffusion, pseudodiffusion (D*) related to intravascular water molecule movement or perfusion, and flow fraction (f) (Le Bihan et al., 1988 ).

DWI has been reported to be a good predictor of renal changes in diabetic kidneys (Deng et al., 2018; Feng et al., 2018), as well as a powerful technique for monitoring post-transplant renal function (Chen et al., 2018; Fan et al., 2019 ; Ren et al., 2016; Xie et al., 2018). Several studies have demonstrated DWI as a promising imaging technique for renal fibrosis assessment in animal models (Cai et al., 2016; Hennedige et al, 2015) and human studies of CKD (Ding et al, 2016; Friedli et al, 2017 ; Gaggioli et al., 2007; Ichikawa et al., 2013; Q. Li et al., 2014; Xu, 2010). ADC values ​​correlate with renal function, usually assessed by creatine levels or eGFR (Ding et al, 2016; Gaggioli et al, 2007; Xu, 2010), as well as renal fibrosis and pathology scores at biopsy (Cai et al, 2016; Ebrahimi et al, 2014a; Friedli et al, 2017; Q. Li et al., 2014; when related to renal function, DWI parameters were found to decrease with severe renal damage (reflected by decreased eGFR), with increased fibrosis associated with decreased adc There is a strong correlation. The decrease in parameters can be attributed to reduced perfusion, the presence of interstitial fibrosis limiting water, and reduced vasculature (Hennedige et al, 2015). In addition, cortical and medullary DWI parameters (-ADC, -D) has a good negative correlation with the percentage of fibrosis (Friedli et al, 2017). The adc group (control group, mild, moderate, severe) based on the glomerular segment hyperplasia score as the pathological grading standard ) are significantly different, and the pathological types (nephropathy, small change glomerulonephritis, focal segmental proliferative glomerulonephritis, membranous nephropathy, mesangial proliferative glomerulonephritis, glomerulosclerosis, There is no significant difference between the adc groups for which the pathological grading standard is crescent glomerulonephritis). This may mean that different pathological types of CKD have similar pathogenic characteristics, leading to a decrease in adc (Q. Li et al. , 2014). However, ADC values ​​failed to detect CKD early as the ability to differentiate between healthy and stage I CKD was not mentioned, whereas perfusion-related D* supported that kidney damage is accompanied by reduced perfusion in one study, and IVIM maps should be better than ADC maps Better early detection of renal dysfunction (Ichikawa et al., 2013). On the other hand, the performance of using ADC to detect renal damage (eGFR <=30 ml/min/1.73 m2 or biopsy confirmed) is better than that of IVIM parameters (Ding et al., 2016; Friedli et al., 2017). Compared with diabetic patients without kidney disease, the average ADC value of patients with advanced diabetic nephropathy (DN), which is the main cause of renal failure, is significantly reduced (Jawad, 2019). Recently, DWI has shown promise in differentiating healthy children from CKD (glomerulonephritis, hemolytic uremic syndrome, lupus nephritis (LN), nephropathy, and infantile nephropathy are the main causes of CKD) (Emad-Eldin et al, 2020 ). adc is negatively correlated with CKD stage (stage I to V) (Emad-Eldin et al, 2020). It was reported that the presence of renal allograft dysfunction associated with fibrotic deposition (mean eGFR 30 ml/min/1.73 m2) significantly reduced cortical and medullary ADCS (Bane et al, 2020). 2020). adc is negatively correlated with CKD stage (stage I to V) (Emad-Eldin et al, 2020). It was reported that the presence of renal allograft dysfunction associated with fibrotic deposition (mean eGFR 30 ml/min/1.73 m2) significantly reduced cortical and medullary ADCS (Bane et al, 2020). 2020). adc is negatively correlated with CKD stage (stage I to V) (Emad-Eldin et al, 2020). It was reported that the presence of renal allograft dysfunction associated with fibrotic deposition (mean eGFR 30 ml/min/1.73 m2) significantly reduced cortical and medullary ADCS (Bane et al, 2020).

Research incorporating DWI into clinical applications remains limited. Most of these studies evaluated the relationship between ADC and fibrosis or pathology scores. Few studies have focused on the use of DWI to differentiate healthy patients from those with CKD or DN and to detect renal allograft dysfunction. Although unlike biopsy, DWI provides the ability to assess whole-kidney perfusion and diffusion, it is still limited to respiratory motion artifact, protocol variability, inter-subject variability (e.g., patient preparation), and the impact of factors other than fibrosis on The influence of water mobility (e.g. urinary flow rate, drugs, vascular volume) makes it impossible to use clear cutoff values ​​for ADC (Gaggioli et al., 2007; Notohamprojo et al., 2010; Sulkowska et al., 2015). Clearly, future studies are needed to address these limitations and further explore the potential of DWI in renal function assessment in CKD patients, as well as comparison of ADC and IVIM classification performance.

2.1.2 Diffusion tensor imaging (DTI)

Renal fibrosis or matrix deposition is expected to not only reduce the mobility of water molecules within renal tissue but also disrupt the substantial orderly structure that facilitates water flow in specific directions (Leung et al, 2017). Diffusion tensor imaging (DTI) is an advanced diffusion magnetic resonance technique that can measure the mobility of water molecules along different axes, providing the opportunity to assess kidney microstructural organization and capture ordered structural disruptions (Leung et al., 2017) . DTI has diffusion measurements in at least six directions from which fractional anisotropy (FA) and mean diffusivity (MD) can be derived. FA represents a measure of diffusion anisotropy, ranging from 0 to 1, as an indicator of the degree of anisotropy or diffusion directionality, while MD is equivalent to the ADC derived from the DWI technique, reflecting the diffusion amplitude (Notohamiprodjo et al, 2010). Furthermore, three-dimensional reconstruction of the diffusion direction can be performed using microbeam imaging techniques that combine direction and anisotropy indices, as shown in Figure 1 (notohamiprojo et al., 2010).

Renal pyramidal FA has been reported to be higher than cortical FA, indicating fiber accumulation, reflecting the arrangement of tubular structures or tubular flow within the usually highly structured medulla (Notohamiprodjo et al., 2010; Kataoka et al., 2009; Nicolaescu et al. , 2017; Notohamprojo et al., 2010; Ries et al., 2001). DTI has been used to evaluate mouse kidney disease models and was found to be a good marker of renal pathology and renal fibrosis (Hueper et al., 2012; J.-Y. Kaimori et al., 2017; j. .y.j.y.Kaimori et al, 2017). Kidney disease is the leading cause of end-stage renal disease, and in predicting kidney disease, DTI is expected to detect kidney infection early even if eGFR remains normal (>60 mL/min/1.73 m2). There is a good correlation between FA and eGFR. Compared with the control group, FA is significantly lower in DN patients (eGFR > or <60 mL/min/1.73 m2) (Lu et al, 2011). Furthermore, medullary FA has been reported as a good biomarker of renal function in allogeneic kidney transplants, and it correlates well with eGFR as well as the amount of renal fibrosis (Hueper et al., 2016; Lanzman et al., 2013; Palmucci et al., 2016). al., 2015). When examined histologically, FA correlates with glomerulosclerosis percentage, interstitial fibrosis area, and degree of renal damage (stages 1 to 5) (Feng et al, 2015).

In CKD, medullary FA is significantly decreased in patients with chronic renal impairment (eGFR <60 mL/min/1.73 m2) (Saini et al, 2018). Renal failure is caused by: renal vascular sclerosis, renal artery stenosis, renal marrow cystic disease, DN, pyelonephritis, stones, acute glomerulonephritis, analgesic abuse, interstitial nephritis, Wegener's granulomatosis, LN, diuretic abuse (Gaudiano et al, 2013)) and normal kidneys Compared with functional controls, it is closely related to eGFR (Gaudiano et al, 2013); Saini et al, 2018; W. Wang et al., 2014). Urethrography also showed good visual differentiation between the two subjects, with a reduced number of urethras without preferential orientation and worsening renal function compared with conventional urethral arrangement with normal renal function (Gaudiano et al, 2013). Ye et al., DTI has been successful in early detection of CKD progression in diabetes mellitus (DM). Compared with controls, cortical and medullary FA were significantly decreased in stage 1 CKD and were well correlated with eGFR, indicating that DTI can predict CKD progression (Ye et al, 2019). However, DTI cannot identify changes in the medullary diffusion direction in the early stages of DN (Feng et al, 2019). It has been reported that medullary FA and mean renal shelf length were significantly reduced in children with autosomal recessive polycystic kidney disease (ARPKD) compared with healthy controls (Serai et al., 2019). Recently, DTI parameters, FA, and track length were significantly different between kidneys with and without ureteropelvic junction (UPJ) obstruction, supporting the ability of DTI to assess parenchymal damage (Otero et al, 2020). In addition, it has been reported that gadolinium-based DTI is more accurate in measuring renal pathological characteristics, renal fibrosis, and renal blood flow in patients with stage I and II CKD (60<=eGFR<=90 mL/min/1.73 m2) (Liu et al., 2020). Renal artery stenosis (RAS) is a kidney disease that promotes fibrosis through collagen deposition. In patients with RAS who have significantly reduced medullary FA, ​​DTI can detect RAS-induced changes in diffusion parameters (Gaudiano et al, 2020). In (Mrd¯anin et al., 2020), medullary FA was found to be lower in DM patients and positively correlated with eGFR. Nephrography shows the distribution structure in patients with renal damage. Cortical FA and medullary FA are related to allograft function, with well-functioning (stages I and II) grafts having significantly higher FA than dysfunctional (stages III-V) grafts (S. et al., 2020). It has been reported that medullary and corticomedullary differentiation (CMD) of FA is closely related to eGFR in healthy and transplanted kidneys (Adams et al., 2020).

DTI is reported to reflect the severity of kidney damage. In summary, the clinical application of DTI aims to discover the relationship between DTI parameters and eGFR or fibrosis, and to evaluate the effectiveness of DTI parameters and nephrography in distinguishing healthy volunteers from patients with renal impairment (DN, DM, UPJ obstruction, RAS ), to evaluate the potential application of DTI in the early detection of DN and DM, and to test the value of DTI in reflecting the function of allogeneic kidney transplants. Medullary FA is the main DTI indicator of renal damage and decreases as renal function worsens. Many factors may be involved in the process of renal damage, which can lead to a decrease in FA, such as reduced tubular flow velocity, tubular damage, and vascular abnormalities (Lu et al, 2011).

2.2 Blood oxygen level dependent imaging

Renal hypoxia has been considered to play an important role in the progression of CKD (Nangaku, 2006). Blood oxygen level-dependent magnetic resonance imaging (BOLD) provides the opportunity to measure tissue oxygenation levels without the use of contrast agents. Briefly, BOLD-MRI exploits the paramagnetic properties of deoxyhemoglobin to assess tissue oxygenation: the higher the local deoxyhemoglobin level, the higher the apparent relaxation rate R2* (s−1) and the lower the local tissue oxygen content.

Some animal studies have shown a linear relationship between R2* values ​​and directly measured renal partial pressure (pO2), which reflects the amount of oxygen dissolved in the blood using oxygen-sensitive electrodes. Therefore, BOLD can accurately measure tissue pO2 (prujm et al, 2018a). Many factors can alter renal oxygenation (R2* value), such as hydration status, sodium balance, and use of medications that may affect oxygen delivery or consumption. Therefore, patient preparation needs to be standardized, including fasting for 4–6 hours at night, constant water intake every hour, controlling salt intake, monitoring medication intake, and recording blood pressure and medications before the MRI examination (Prujm et al. , 2018a), (Neugarten, 2012).

To analyze the BOLD images, four methods were used: the Region of Interest (ROI) technique, the Twelve Layer Concentric Object (TLCO) or "onion skin" technique, fractional tissue hypoxia, and the compartmental method, as shown in Figure 2. As shown in the figure, the ROI technique requires the placement of different small ROIs (20-40 voxels) in the cortical and medullary regions to provide separate information. The TLCO technique divides the renal parenchyma into 12 layers of equal thickness, and calculates the average R2*. From artificially selected renal parenchyma, fractional tissue hypoxia is reported as the percentage of R2* values ​​above a certain threshold. Compartmental method was used to analyze the distribution of R2* values ​​in the renal region.

The hypoxia hypothesis suggests that renal tissue hypoxia or reduced tissue oxygenation is the final common pathway of CKD. Initially, studies failed to find a relationship between R2* values ​​and eGFR-estimated kidney function (Michaely et al, 2012; prujm et al, 2014). Even, few studies have reported decreased or no changes in cortical R2* associated with renal impairment (Djamali et al., 2007; Khatir et al., 2015; Textor et al., 2008; Wang Zhijun et al., 2011). In contrast, recent studies have shown a good correlation between cortical R2* or R2* slope and eGFR, with R2 in patients with renal impairment (glomerulonephritis, hypertensive nephropathy, DN, acute kidney injury (AKI), etc.) The ∗ value is large and limited to the cortex, supporting the hypoxia hypothesis (Inoue et al., 2011; Milani et al., 2017; Prasad et al., 2018; Project et al., 2018b; Xinlong et al., 2012; Yin et al., 2012) . This difference between the results could refer to the fact that R2* is not only lacking or even dependent on intravascular deoxyhemoglobin, but also depends on other factors such as renal edema and perfusion, changes in vascular and tubular volume fraction (Milani et al., 2017; Prujm et al. , 2017).

Milani et al showed that the BOLD image analysis technique can influence the results between controls and CKD patients (eGFR<=60 mL/min/1.73 m2 or proteinuria <300 mg/24 h for at least 3 months) using ROI technique showed no change in R2* between patients, whereas the R2* slope using the TLCO technique was significantly different from the same group of patients (Milani et al., 2017). Likewise, area R2* shows a strong correlation between ROI and TLCO technology. A. Li et al., 2020). However, the TLCO parameter R2* slope showed the greatest sensitivity in terms of distinguishing controls from CKD patients. The association between adc and R2* demonstrated by Prasad et al., and the finding that R2* increases with fibrosis stage by Inoue et al., support the hypothesis that fibrosis is a way to reduce local oxygen consumption (Inoue et al. et al., 2011; Prasad et al., 2018). The assessment of fibrosis by BOLD imaging has also been validated in animal models, where R2* values ​​correlate well with fibrosis percentage (Woo et al, 2018; Zha et al, 2019). In the study by prujm et al. (prujm et al., 2014), the use of urea to induce an increase in R2* values ​​improved the differentiation of healthy and damaged kidneys (prujm et al., 2014), but no positive effects were subsequently found (Milani et al. People, 2017; Prujm et al, 2018b). Notably, cortical R2* values ​​are closely related to fluid fraction (FF) compared with GFR, and in healthy volunteers FF = GFR/renal plasma flow. Therefore, it is recommended that FF be measured in future studies (Van Der Bel et al, 2016). It has been reported that decreased oxygenation reflected by elevated R2* values ​​is a clinically useful marker of CKD progression (Sugiyama et al., 2018). In their study, controls and diabetic patients were included in a retrospective study over 5 years, in which the rate of eGFR decline was significantly correlated with the R2* value. Medulla R2* value has been shown to be sensitive to the prediction of early DN (Feng et al, 2019). In addition, it has been reported that medullary R2* can distinguish patients with mild renal impairment (CKD stage I, stage II) with high sensitivity and specificity (92.3% and 85. 2%) (Li et al, 2019). Incorporating R2* values ​​into machine learning algorithms was able to differentiate between rejected and non-rejected transplants with 100% sensitivity (Shehata et al., 2019). Recently, cortical and medullary R2* values ​​were positively correlated with serum creatinine levels and negatively correlated with eGFR (Luo et al, 2020). In their study, cortical R2* values ​​were classified as CKD stage IV, V > CKD stages I-III > healthy controls, medullary R2* values ​​were classified as CKD stage IV, V > CKD stages I-III, and There were no significant differences between controls and patients with CKD stages I-III.

Clinical applications of MRI imaging studies using BOLD are limited to studying the relationship between R2* values ​​and eGFR, serum creatinine levels, fibrosis, and FF, evaluating the potential use of BOLD in differentiating controls from patients with renal impairment, and evaluating R2* values ​​in early To examine the role of DN and discover the value of BOLD in reflecting renal allograft function. Considering that the R2* slope derived from the TLCO technique is related to some extent to the R2* distribution within the renal parenchyma, the door opens to further explore the spatial arrangement within BOLD. The usefulness of R2* in staging CKD or predicting renal function decline and measuring fibrosis requires more follow-up studies.

2.3 Arterial spin labeling

Since CKD is associated with impaired perfusion and reduced renal blood flow due to renal fibrosis, assessment of renal perfusion may be a powerful tool for detecting CKD and distinguishing several CKD stages (Leung et al., 2017; Nangaku, 2006). Arterial spin labeling (ASL), in MRI mode, allows tissue perfusion measurements without the administration of gadolinium, using water in the blood as a contrast agent. Blood flow is marked with the opposite magnetization to the target tissue. The difference between labeled and unlabeled images can provide perfusion-weighted images with signal intensity proportional to perfusion (Artz et al., 2011; Leung et al., 2017; Odudu et al., 2018). The signal of each voxel is then input into the kinetic model to generate a perfusion map, as shown in Figure 3.

Several studies support the feasibility of using ASL to assess renal function in patients with CKD or kidney transplantation. Artz et al. found a good correlation between native kidney and transplanted kidney cortical perfusion and eGFR (P<0.05), supporting that glomerular filtration rate regulates renal blood flow. When eGFR >60, the blood perfusion of the transplanted kidney is significantly lower than that of the native kidney, which may indicate that there are differences in the regulation of blood flow by the transplanted kidney, and there is no statistical difference in blood perfusion between the left and right kidneys (Artz et al, 2011).

Applied to patients with renal impairment, cortical and whole-renal ASL perfusion in CKD patients (staged from I to V according to eGFR) is significantly lower than in healthy subjects and decreases with increasing renal impairment (Brown et al., 2019 Gillis, 2016; Rossi et al., 2012), with and without age and body mass index matching (L. Li et al., 2017). Cortical and whole-renal perfusion are well correlated with eGFR (Gillis, 2016; Li et al, 2017). There was no correlation between kidney volume or size and perfusion, suggesting that differences in tissue perfusion cannot be attributed to tissue atrophy (Gillis, 2016). The filtering fraction (FF=eGFR/ASL) determined from ASL perfusion was found to correlate with the fibrosis score from biopsy results (Brown et al, 2019). It has been reported that cortical perfusion is closely related to renal function in healthy and CKD patients (15<=eGFR<60 mL/min/1.73 m2) (Buchanan et al, 2019). Recently, ASL values ​​were found to be significantly different between patients with primary glomerular disease (PGD) and healthy volunteers, and were well correlated with eGFR in PGD patients (CA Li et al., 2020).

ASL has been used in the study of AKI (Dong et al., 2013), LN (Rapacchi et al., 2015; Skeoch et al, 2017), CKD (Breidthardt et al., 2015; Gillis et al., 2016; L. -P.L.-P. Li et al., 2017), DN (Moraguti et al., 2017) and renal transplantation (Heusch et al., 2014; Hueper et al, 2015). From these studies, consistent results are that cortical perfusion decreases with the presence of renal damage, decreases with increasing CKD stage, and correlates with eGFR and fibrosis stage. Although it is capable of assessing renal dysfunction, in order to further explore the relationship between reduced renal perfusion and ECM accumulation, more validation is needed to assess the relationship between perfusion and histopathology.

2.4 Dynamic contrast-enhanced magnetic resonance imaging

Dynamic contrast-enhanced Magnetic Resonance Imaging (DCE-MRI), also known as MR renography, is a magnetic resonance perfusion technology that uses contrast agent to enter the patient's blood. Precipitating fluid during the reproduction of flushing catalyzes differential imaging of the organ (Derle and Dighe, 2015). The interaction between gadolinium-based agents and water protons within the target tissue results in a reduction in T1 relaxation time, resulting in higher intensity on T1-weighted images (Ebrahimi et al., 2014b). To track the dynamic enhancement of the target tissue, consecutive images are acquired during the transition of the contrast agent. Tracer concentration is estimated from MR signal changes using a mathematical model, as shown in Figure 4 (Zhang and Lee, 2019), (Octavia et al, 2017). By fitting the model to the concentration-time curve in the kidney, perfusion and blood flow parameters can be estimated, including renal cortical and medullary plasma flow, plasma volume, mean transit time per compartment and throughout the kidney, and tubular flow and tubular transit time, most importantly single-kidney GFR (SK-GFR) rather than bilateral GFR estimated by serum creatinine levels (E. Eikefjord et al., 2015; Jiang et al., 2019b; Octavia et al, 2017). DCE-MRI requires appropriate segmentation and registration processes for kidney motion correction (Derle and Dighe, 2015).

Despite its ability to visualize kidney status and estimate functional kidney parameters, as well as its prognostic and diagnostic potential in living kidney donors (Dujardin et al., 2005; Eikefjord et al., 2016; E. 2015; Notohamiprodjo et al., 2011), The application of renal DCE-MRI in patients with CKD is limited by the use of contrast agents and the risk of nephrogenic systemic fibrosis even at low doses, resulting in concerns about the safety of renal damage in patients (Ebrahimi et al., 2014b), ( Zhang and Lee, 2019), (Fraum et al., 2017). Therefore, ASL technology is used to assess renal perfusion in patients with renal insufficiency (Zhang and Lee, 2019), (Conlin et al, 2017).

2.5 T1 and T2 mapping

Magnetic Resonance Relaxometry (MRR) can distinguish tissue composition by measuring T1 (spin lattice) and T2 (spin) relaxation times, by performing pixelated mapping of the true T1 and T2 values ​​of the target tissue, and No contrast agent is required. It is used to non-invasively characterize kidney tissue.

T1 relaxation time increases with the progression of tissue scarring. It correlates well with biopsy-measured cardiac fibrosis and is significantly higher in patients with more severe fibrosis stages (Bull et al, 2013), thus the potential of T1 maps to identify fibrotic burden in the absence of contrast agent . Furthermore, T1 values ​​were found to be modulated by tissue and/or blood oxygenation levels in some studies, which demonstrated the sensitivity of cortical T1 values ​​to changes in oxygenation levels (higher oxygenation levels, lower T1 values) (Wolf et al. al, 2018). Furthermore, T1 native localization showed good to strong inter- and intra-exam reproducibility in both healthy and diabetic nephropathy patients (Dekkers et al., 2019).

Renal T1 mapping is used as a non-invasive tool to assess kidney function. In their study, Lee et al. reported that cortical T1 relaxation time is longer than that of the medulla, which is due to higher water mobility within the medulla than the cortex. Cortical T1 increases with renal insufficiency, which can be attributed to There is a good correlation between longer-term pathological changes (such as extracellular matrix changes) leading to increased cortical water content and cortical T1. Difference between cortical and medullary T1s and single kidney GFR (SK-GFR) (r= - 0.5, P = 0.03; r = 0.58, P<0.01) (Lee et al, 2007). Breidthardt et al evaluated the relationship between renal dysfunction, perfusion (ASL) and true parenchymal structure (T1 map) in healthy volunteers and heart failure (HF) patients with varying degrees of renal impairment based on eGFR (Breidthardt et al, 2015b) . The main results showed that T1s had a good correlation with eGFR (r = 0.41, P = 0.015), and cortical T1s was higher in HF patients with renal insufficiency than in HF patients without renal insufficiency. Similar results were obtained in the study of Gillis et al. In clinically evaluated CKD patients, cortical t1 levels were higher than normal subjects and correlated with eGFR (r = 0.58, P<0.001) (Gillis, 2016). Furthermore, Friedli et al (Friedli et al., 2016) demonstrated the potential application of T1 mapping in renal fibrosis assessment, where CMD of T1 was positively correlated with eGFR and renal fibrosis stage, whereas cortical or medullary T1 failed to show correlation, which may be explained by the ability of the CMD technique to reduce inter-individual variability resulting from the selection of ROIs from the cortex and medulla respectively. Cox et al. performed a multiparametric study of renal perfusion, diffusion, oxygenation, and microstructure in patients with CKD using ASL, DWI, T2*, and T1 mapping, and they found that cortical T1s were significantly increased, whereas CMD of T1s increased in renal impairment ( CDK stages III and IV) (Cox et al, 2017). Furthermore, T1 native mapping has been studied in kidney transplant studies to assess kidney function, where T1s showed an increase in transplanted kidneys compared with native kidneys, an increase in damaged kidneys compared with well-functioning kidneys, and a decrease in CMD of T1s was found after transplantation (Huang et al., 2011;Peperhove et al,

Research on renal T2 measurement remains sparse. Cortical T2s were significantly increased in transplanted kidneys compared with native kidneys, while there was no correlation between T2s and allograft function (eGFR) (Mathys et al., 2011). Furthermore, T2 localization has been applied to autosomal dominant polycystic kidney disease (ADPKD), which is characterized by the progressive development of cystic and fibrotic components, and whole-kidney T2s were found in animal models to distinguish healthy kidneys from ADPKD kidneys (Franke et al., 2017). Recently, it was reported that T1 and T2 were able to assess cystic kidney disease progression in kidneys with ARPKD in a mouse model, with T1 and T2 being significantly higher in affected kidneys compared with healthy kidneys (MacAskill et al., 2020 . Interestingly, T2s were found to be closely associated with cystic scores in ADPKD patients (sidek et al, 2020).

In summary, T1 images demonstrate the ability to assess renal damage, whether that is reflected by low eGFR, HF associated with renal dysfunction, or renal transplantation. A consistent finding is that t15 correlates with GFR and fibrosis stage, increasing when renal damage is present. The CMD of T1s has been reported to decrease with renal dysfunction and was found to be more effective than cortical or medullary T1s alone, providing an opportunity to further study the distribution of T1s within the intact renal parenchyma. In clinical studies of polycystic kidney disease (ADPKD and ARPKD), T1 and T2 profiles were introduced to a small extent, where T1s and T2s were found to be elevated in affected kidneys, and T2s was reported to have a good correlation with renal cystic fraction. Additionally, the T2 map is not widely explored. Therefore, further studies are needed to evaluate their potential in the assessment of renal dysfunction. Finally, further verification is needed of the correlation between relaxation mapping and renal function reflected by GFR, as well as the sensitivity of the mapping to detect fibrosis.

2.6 Magnetization transfer magnetic resonance imaging

Magnetization transfer magnetic resonance imaging (MT) has the potential to assess fibrosis in biological tissues. MT is sensitive to fixed macromolecular components in tissues and can assess pathological events accompanied by macromolecular changes (such as fibrotic components). It is sensitive to the interaction between free and fixed protons in macromolecules. Using nonresonant radiofrequency pulses, macromolecular protons are saturated and transferred to water protons based on the exchange rate between the two proton populations. Two sets of images need to be acquired, one is the baseline image without MT pulse, and the other is the MT-weighted image. Therefore, the percentage reduction in water signal after exchange reflects the MT ratio (MTR), which is an indication of macromolecule content (Henkelman et al, 2001).

MT has shown its utility in detecting fibrosis in animal models and in patients with kidney disease. In a mouse model of RAS, Ebrahimi et al. demonstrated the ability of MT to visually differentiate between fibrotic and non-fibrotic tissue, where MT-derived parameters (magnetization fractional pool and free magnetization exchange rate) varied between the stenotic and contralateral kidneys. were significantly different and correlated with fibrosis quantified by tricolor tissue staining. Therefore, they have the potential to serve as biomarkers of renal morphological changes following the composition of fibrosis (Ebrahimi et al, 2013). Similarly, in a mouse model of RAS, MT successfully monitored renal fibrosis longitudinally, with median MTR in the cortex and medulla significantly reduced in RAS kidneys and strongly associated with anterior fibrosis as assessed by histology. Furthermore, Trichomes and Sirius red staining showed good spatial consistency between MTR patterns and renal fibrosis (Jiang et al, 2017a), (Jiang et al, 2017b). MT was tested on a mouse ADPKD model as it is accompanied by cystic (reduced MTR) and fibrotic burden (increased MTR). The parameters of the MTR chart (including mean, median, 25th percentile, skewness and kurtosis) are closely related to renal pathological indicators. MTR has a good correlation with cystic and fibrosis histological indicators. Histological staining films The consistency with the MTR chart is good, as shown in Figure 5 (Kline et al, 2016). Recently, MTR has been shown to provide structural and metabolic assessment of renal fibrosis in rats with unilateral ureteral obstruction (UUO), where MTR was significantly different between healthy kidneys and contralateral ureteral obstruction kidneys during obstruction, and closely related to metabolic markers (aali et al., 2020).

Since MTR is affected by many factors related to sequence details and relaxation parameters, quantitative MT (qMT) technology has been developed to provide a more quantitative assessment of intra-tissue macromolecule content with higher sensitivity and specificity (ff Wang et al ., 2018). In this technique, the ratio of the macromolecular proton pool to the free pool (pool ratio size, PSR) is separated from the relaxation rate and exchange rate and used as a quantitative MT parameter. Using qMT, fibrosis was assessed in mice with progressive diabetic nephropathy and tubulointerstitial fibrosis, and PSR has been shown to be a useful indicator of fibrosis (ff Wang et al., 2018), (Wang et al., 2019 ).

Additionally, a human study found a role for MT in assessing renal fibrosis. Cortical MTR correlates with eGFR and differs significantly between subjects with normal renal function and patients with renal impairment, classified according to their eGFR (Ito et al., 2013).

MTR and PSR have proven to be promising biomarkers of fibrosis. As a common pathway of CKD, the ability of MT to detect fibrosis has been demonstrated in animal models, but its clinical application in human studies, especially in the identification of CKD, remains limited.

2.7 Magnetic resonance elastography

Organ stiffness has been shown to precede fibrosis and extracellular matrix deposition (Georges et al, 2007). Fibrosis reduces tissue elasticity; therefore, tissue elasticity measurements can provide a good marker of fibrosis (Hewadikaram et al, 2018). Magnetic resonance elastography (MRE) can measure the mechanical properties of tissues by applying mechanical vibrations to the target organ and capturing the shear waves generated and propagated by motion synchronization or phase contrast imaging (Muthupillai et al., 1995). More fast waves with longer wavelengths are captured from stiffer or fibrotic tissue than from soft, healthy tissue. These waves are processed to produce a stiffness map or elasticity map (Figure 6).

MRE has been reported to be a good detector of liver fibrosis in animal and human studies, where increasing stiffness with increasing fibrosis severity was captured, and good predictive power for fibrosis stage was found. In terms of fibrosis staging, MRE even surpassed ADC atlas with better sensitivity and specificity (Kim et al., 2013; Rouvi et al., 2006; Talwalkar et al., 2008; Wang Yuyu et al., 2011; Yin et al., 2007b , 2007a).

Likewise, changes in renal mechanical properties, elasticity, and stiffness can also be detected by MRE and found to be associated with renal fibrosis. Shah et al. demonstrated the ability of MRE to measure renal cortical stiffness induced by mild fibrotic nephrocalcinosis in a mouse model (Shah et al, 2004). A significant increase in medullary stiffness was found in a porcine model of RAS and was closely related to the degree of fibrosis assessed histologically (Korsmo et al., 2013; Zhang Xu et al., 2018). Lee et al reported a modest increase in stiffness in fibrotic kidneys (Lee et al., 2012). Furthermore, renal stiffness correlates with eGFR and is significantly higher in functional compared with non-functioning transplanted kidneys (Garcia et al., 2016). In a follow-up study, it was found that whole-kidney stiffness was associated with fibrosis scores and eGFR in kidney transplant patients, and eGFR increased as eGFR decreased in kidney transplant patients (Kim et al., 2017; Kirpalani et al, 2017). Similarly, there is a negative correlation between eGFR and MRE stiffness (Zhang and Zhang, 2020). Recently, Hodneland et al. reported a good correlation between shear wave-derived parameters (reflecting pressure gradient, volume, and shear deformation) and biopsy-determined grade of arteriosclerosis (Hodneland et al., 2019). Surprisingly, patients with all stages of CKD DN (stages I to V) had reduced stiffness compared with controls, which could be affected by hemodynamics (blood flow) and fibrotic deposition via MRE. to explain the impact (Brown et al, 2019).

MRE has been reported to be an excellent tool for fibrosis detection and staging and can serve as a predictor of kidney function. The main results showed that increased stiffness was associated with increased fibrosis stage. Clinical applications in renal transplantation have reported a correlation between MRE stiffness and eGFR, with increased stiffness associated with the presence of renal dysfunction. However, renal stiffness can be affected by factors other than fibrosis, such as hydronephrosis, edema formation, renal blood flow, collecting system dilation, changes in paramagnetic material composition (e.g., lipids, proteins) and structural factors (various anisotropic structures), these factors can mask fibrosis during MRE interpretation (Leung et al., 2017), (Lee et al., 2012; Warner et al., 2011), (Gennisson et al., 2012). These results require further studies to test its ability to diagnose patients with CKD.

2.8 Other magnetic resonance imaging techniques

Sensitivity-weighted imaging (SWI) shows promise in renal function assessment and fibrosis detection. Mie et al studied the feasibility of SWI on human kidneys (Mie et al, 2010). SWI signals were found to be affected by perfusion changes and tissue fibrosis in animal models (Pan et al., 2017; Zhang Jiangang et al., 2018). Pan et al demonstrated the sensitivity of SWI after renal reperfusion, with SWI scores decreasing and returning to baseline levels within 48 hours after reperfusion injury (Pan et al., 2017). In addition, Zhang et al.'s study showed that the SWI signal ratio is reduced and has a strong correlation with the stage of fibrosis (JGJG Zhang et al, 2018).

In addition, fat quantification using the Dixon technique (also known as fat fraction imaging) has been investigated for early detection of renal lipid deposition in patients with diabetic nephropathy, with a decrease in fat fraction noted to correlate with the presence of the disease (y.c.). Wang et al., 2018). The use of the Dixon technique in the assessment of CKD deserves further study to examine its ability to detect kidney damage.

In addition, magnetic resonance spectroscopic imaging (MRSI), which characterizes metabolites within tissues, has been investigated for the assessment of kidney disease. Given that renal failure is associated with the progressive formation of inorganic phosphorus and the loss of adenosine triphosphate, 31P magnetic resonance spectroscopy has emerged using the ratio of phosphorus homoesters to inorganic phosphorus as a marker of renal metabolism (Ebrahimi et al., 2014b). The metabolomic signature successfully reflects different renal allograft functions as it is closely related to eGFR (Bassi et al., 2017). CKD metabolic biomarkers are markers of glomerular filtration, tubular and mitochondrial function, urea cycle, or amino acid changes that disappear with increasing severity of injury (Hocher and Adamski, 2017).

3 Other imaging methods

3.1 Ultrasound elastography (UE)

Ultrasound elastography (UE) can detect the mechanical properties of tissue, such as MRE. UE techniques are based either on imaging of generated shear waves propagating within the tissue or on strain analysis of the tissue under external compression (Gennisson et al, 2013). Shear wave elastography (SWE or USE) uses sound waves to measure shear wave velocity (SWV), which reflects tissue stiffness, to assess tissue elasticity. Strain elastography (SE) applies external stress to the kidney and measures the deformation due to this stress, known as cortical or medullary strain and the cortical-medullary strain ratio. Compared with traditional Doppler ultrasound (US)-derived parameters, ue-derived parameters have shown superior potential in assessing renal fibrosis (Hu et al, 2015; Leong et al, 2019; Marticorena Garcia et al, 2018).

Since the kidneys are close enough to the body surface, SE is used to evaluate renal function after transplantation. The cortical-medullary strain ratio shows a downward trend with the increase of cortical fibrosis degree, and there are significant differences between different fibrosis score groups, and is negatively correlated with fibrosis stage (Gao et al., 2015, 2013a). When normalized, cortical strain was found to have good discrimination between mild and moderate fibrosis stages (Gao et al., 2013). In addition, the average tissue elasticity is negatively correlated with the degree of fibrosis (Orlacchio et al., 2014). Garcia et al used SWE to detect renal function in allograft patients and reported that functional kidneys were stiffer than non-functional kidneys with good predictive performance (sensitivity 90.9%, specificity 85.7%, AUC of pyramid stiffness 0.925) , SWV correlates well with renal blood flow and eGFR (Marticorena Garcia et al., 2018). Furthermore, some studies reported a good correlation between SWE hardness and fibrosis stage (Arndt et al., 2010; Ma et al., 2018; Nakao et al., 2015). In contrast, some studies reported that SWV or estimated stiffness was not correlated with renal function and did not differ among grafts at different stages of fibrosis (Grenier et al, 2012; Lee et al, 2015; Syversveen et al, 2012, 2011).

Applying SWV to CKD patients, it was found that SWV has a good correlation with eGFR. Compared with various stages of CKD patients, SWV in healthy patients was significantly higher (AUC=0.752), but it cannot distinguish CKD stages (CKD based on eGFR staging) (Guo et al., 2013), (ll Wang et al., 2014). It has been reported that the decrease in SWV is related to the severity of renal damage or histological assessment of fibrosis score (Hu et al., 2015). Compared with conventional ultrasound parameters, including kidney length, parenchymal thickness, and resistive index, SWV exhibits better differentiation performance (Hu et al., 2015). Similarly, Bob et al. also reported that a decrease in SWV was associated with a decrease in eGFR-defined renal function in patients with diabetic kidney disease (DKD) without other renal disease or DM (Bob et al., 2017). The SE strain ratio was found to have a good correlation with eGFR in DN patients (Iacob et al, 2019). On the other hand, no correlation was found between biopsy-assessed renal stiffness and fibrosis scores (Cardenas et al., 2019), (LL Wang et al., 2014), and estimates were found in CKD patients (stages III to V) Stiffness is higher (Lin et al., 2017; Samir et al., 2015). Recently, SWE was found to be able to detect abnormal renal stiffness in patients with early glomerulonephritis and preserved renal function (Grossmann et al, 2019). Young's Modulus (YM) was used to assess kidney stiffness and showed the greatest discriminatory power between healthy volunteers, type 2 diabetic patients without DKD, and type 2 diabetic patients with DKD (Shi et al. al., 2020). There is a negative correlation between YM and eGFR, and the eGFR of healthy controls is higher than that of patients. SWE outperforms traditional US parameters and shows its ability to monitor type 2 diabetes (Shi et al., 2020).

UE has not yet shown the ability to objectify the staging of renal impairment. In terms of clinical application, SE is used in kidney transplantation research, and the strain ratio or average tissue elasticity has a good correlation with fibrosis stage. Functional allografts have higher stiffness than nonfunctional allografts. SWE has been included in studies of CKD, DKD, glomerulonephritis, DM, and DN, leading to conflicting results regarding the relationship between estimated stiffness and renal function. Furthermore, shear waves should propagate faster in fibrotic tissue compared to healthy tissue, which was not noticed in the above studies. The previously reported association between SWV and renal blood flow (Marticorena Garcia et al., 2018) supports the fact that many factors can influence tissue stiffness and mask fibrosis, leading to questions regarding the potential of UE in fibrosis assessment. Conflicting results (Leung et al, 2017), (Lee et al, 2012; Warner et al, 2011), (Gennisson et al, 2012). Interestingly, the spectral parameters obtained from ultrasound should reflect the frequency content and then the disease severity of CKD (Hewadikaram et al, 2018).

3.2 Computed tomography (CT)

Computed tomography (CT) is an imaging technique that uses a motorized x-ray source to emit narrow beams of x-rays and produce tomographic images. CT provides good spatiotemporal resolution and quantitative capabilities for contrast agents (Zhu et al., 2018).

Volumetric CT has shown that renal function of kidney donors can be assessed by measuring SK-GFR (Gaillard et al., 2017; Jiang et al., 2019a; Patankar et al., 2014; Yanishi et al., 2015). Split parenchymal volume correlates closely with schistozoal GFR measured by scintigraphy and is a better indicator of reduced schismatic renal function when combined with schistozoal ADC (Li et al., 2018; Mitsui et al, 2018) . Furthermore, cortical volume has been shown to be a powerful tool for renal function assessment and renal prognosis in kidney donors and a good predictor of CKD development after nephrectomy (Gardan et al., 2018; You et al., 2018). Most importantly, CT shows promise in the assessment of renal function decline and the detection of fibrosis. Since CKD is often associated with renal microvascular rarefaction, Stillfried et al. have demonstrated the potential of CT angiography-derived parameters in the assessment of renal function in CKD, where renal relative blood flow (rBF) closely reflects biopsy-induced renal rarefaction and found that in patients with CKD The arterial diameter was significantly reduced (eGFR <=32 mL/min/1.73 m2) (von Stillfried et al., 2016). In addition, Zhu et al. studied a novel gold nanoparticle conjugated with anti-collagen-i antibody as a CT contrast agent capable of displaying mouse renal fibers by multi-detector CT (MDCT) or micro-CT matching renal histology. ation (Zhu et al., 2018).

Despite the ability of CT to measure SK-GFR and the potential of CT angiography in the assessment of renal function in CKD patients, CT is still limited to ionizing radiation and contrast agent injection that can damage the kidneys (Lerman et al., 1996; Maioli et al., 2012 ). Recently, it has been reported that contrast-enhanced CT increases the risk of end-stage renal disease (ESRD) (Lim et al, 2020). However, identifying renal fibrosis or perfusion through non-contrast CT remains challenging (Zhu et al., 2018).

3.3 Scintillation imaging (PET, SPECT)

Positron emission tomography (PET) and single-photon emission computed tomography (SPECT), also known as scintigraphy, are the most commonly used imaging modalities in nuclear medicine and use radioactive tracers to assess organ function and perfusion. When applying positron-emitting nuclides, PET relies on the detection of radiation emitted through electron-positron annihilation, while SPECT measures gamma rays emitted from the tracer (Köhnke et al., 2019).

Nuclear imaging allows split kidney function measurement (SKGFR) (Patankar et al., 2014; SHIMIZU et al., 2016; Yanishi et al., 2015). Furthermore, PET and SPECT have been shown to be able to detect liver, heart, and pulmonary fibrosis (dsamsog et al., 2017; Kim et al., 2016; Li et al., 2011). In addition, scintigraphy has successfully predicted early post-transplant renal function decline (Yazici et al., 2015, 2013; Yoon et al., 2016), and when combined with CT, the predictive power has improved (Lovinfosse et al., 2016). Recently, nuclear medicine has been shown to be a non-invasive tool for quantitative assessment of glomerular function (Qin et al., 2019). PET has shown potential predictive ability for renal recovery in kidney transplant AKI patients (Pajenda et al., 2020). The results suggest that more research is needed on the role of nuclear imaging in the assessment of renal fibrosis and perfusion in CKD, as clinical applications are limited to SK-GFR measurement, fibrosis detection, and prediction of kidney transplant function.

4 Application of artificial intelligence in medical imaging

In addition to the qualitative analysis provided by the average value of grayscale pixels within a region of interest, texture analysis from mathematical techniques has the potential to provide quantitative information that is typically imperceptible to radiologists. Texture analysis technology characterizes the heterogeneity of pixel distribution and its spatial arrangement by describing the interrelationships and gray-level frequencies between pixels in the image. To quantify image texture, several methods have been involved, including histogram analysis, two-dimensional Fourier transform, statistical methods ranging from gray-level co-occurrence matrices (GLCM), gray-level run length matrices (GLRLM), and local binary patterns ( LBP) derived features), model-based methods (autoregressive and fractal models) and transformation-based methods (wavelet transform) (Larroza et al., 2016). The process starts with image acquisition and involves several steps, typically including ROI definition and preprocessing, feature extraction, feature selection based on the statistical significance of parameters, and classification using simple statistical models or machine learning techniques (Larroza et al, 2016).

Machine learning refers to computer algorithms that learn from observations to predict future outcomes. There is a wide variety of classifiers used in the field of medical imaging, ranging from traditional statistical methods to more complex algorithms. The classification performance of any model uses many measures, such as confusion matrix measures (such as sensitivity, specificity, precision, and accuracy) or the area under the receiver operating characteristic curve (AUC). In traditional machine learning, training and test data should share the same distributed feature space as the real-world input data. The success of predictions largely depends on whether they are true (Weiss et al, 2016). Traditional classifiers can be based on instances (K nearest neighbors), statistical learning theory (support vector machines), decision trees (random forests), feature combinations (linear discriminant analysis, linear or polynomial regression) or probability and statistics (Bayesian) ( Ohata et al., 2019), while deep learning techniques rely on biological neuron structures, which are characterized by nodes consisting of multiple artificial layers related to each other through weights that represent each node's contribution to the output.

4.1 Texture analysis and traditional machine learning techniques

Textures extracted from medical images reflect the micro and macro structures of selected organs (Materka, 2004). Texture analysis is widely used in different medical imaging modalities and has shown its ability as a powerful computer-aided diagnostic tool to aid clinical decision-making (e.g., identification of liver lesions, classification of breast tumors, neoadjuvant breast cancer Prediction of non-response to chemotherapy, classification of renal tumors between simple cysts, kidney stones and complex renal cell carcinoma on CT and MR images) and segmentation tasks such as kidney segmentation based on CT and ultrasound (US) images (Timothy L. Timothy L. Kline et al., 2017; Lubner et al., 2016; Mayerhoefer et al., 2010; Michoux et al., 2015; Raman et al., 2015; Sreelatha and Ezhilarasi, 2018; Yu et al., 2017).

Quantification of tissue fibrosis based on structural analysis has recently emerged. Texture parameters extracted from CT and MR images showed a good correlation between liver fibrosis and increased heterogeneity (Daginawala et al., 2016; Yu et al., 2015; Zhang et al., 2015). Since fibrosis plays an important role in the progression of CKD and considering the heterogeneity of the renal parenchyma (cortex and medulla), texture analysis of renal medical images may be a good predictor of renal dysfunction .

In terms of renal function evaluation, MR, US and scintillation imaging were used for texture analysis, and the results are shown in Table 2. Combined with machine learning techniques, textures are proving to be a great complementary tool at the service of physicians, especially in detecting CKD progression early in the disease. Research using texture analysis focuses on a variety of clinical applications, including early detection of renal dysfunction by differentiating healthy volunteers from patients with mild or non-severe renal impairment, differentiating healthy kidneys from diseased kidneys (CKD, LN, nephropathy), and assessing Relationship between texture and fibrotic component and eGFR, distinguishing rejected from non-rejecting renal allografts, predicting CKD progression in ADPKD patients, and most importantly predicting the five stages of CKD.

Applying texture to different functional magnetic resonance images has been reported in recent animal studies (Zha et al., 2019) and human studies (Rossi et al., 2012), (Alnazer et al., 2019; Ding et al., 2019; Timothy L . Timothy L. Kline et al., 2017; Kociołek et al., 2019; Shi et al., 2018). Histogram- and GLCM-based parameters in renal tissue were extracted on three MRI sequences: DWI, BOLD and SWI (Ding et al, 2019). Derived features correlate well with eGFR. Textures of BOLD and SWI were able to differentiate between control and non-severe renal dysfunction groups, suggesting that textures are capable of detecting renal failure in the early stages of the disease when eGFR cannot detect it (Ding et al., 2019). Consistent with these recently published results, our preliminary study performed texture analysis on DWI MR images and confirmed that texture is affected by CKD (Alnazer et al, 2019). Despite the small sample size, significant differences were found between the two groups (controls and CKD patients) in wavelet-based parameters as well as GLCM-based parameters extracted from renal parenchyma. In patients with LN renal scars who ended up with CKD, glcm-based parameters successfully detected changes in BOLD MRI texture features and had good prediction rates for renal pathology patterns (Shi et al, 2018). The authors have demonstrated that LN injury leads to renal histopathological changes, leading to a decrease in cortical R2* values ​​and affecting normal oxygenation distribution within the kidney (Shi et al., 2018). In UUO animal models, histogram features extracted from R2* plots were able to differentiate the degree of induced fibrosis. In short-term fibrosis assessment, cortical texture decreased significantly and correlated closely with fibrosis percentage (Zha et al, 2019). Furthermore, in a retrospective study of patients with ADPKD, structural analysis of T2-weighted MRI (T2W) was performed to predict renal dysfunction (Timothy L. Timothy L. Kline et al., 2017). Incorporating stable texture features (closely related to eGFR changes) into traditional models (age, eGFR, total kidney volume (TKV)) improves the predictive power of CKD progression. Therefore, structural analysis provides additional insights into existing TKV biomarkers for the prognosis of reduced renal function in ADPKD (Timothy L Timothy L Kline et al., 2017). Kociolek et al. have shown that texturing of DCE-MR images expands its possibilities by adding new information about kidney function (Kociołek et al., 2019). Perfusion map histogram analysis (ASL) confirmed the importance of regional assessment of renal perfusion (Rossi et al., 2012). The authors have demonstrated that CKD is associated not only with changes in perfusion mean values ​​but also with changes in the distribution of perfusion values ​​within the cortex and renal parenchyma (Rossi et al, 2012).

Furthermore, fibrotic deposition and renal damage associated with CKD were assessed through texture analysis of ultrasound (US) kidney images (Ardakani et al, 2017; Chen et al, 2019; Iqbal et al, 2017; Sharma and Virmani, 2017). Textures based on Fourier transforms reflecting spatial frequencies successfully differentiated CKD and healthy kidneys, whereas glcm-based parameters failed (Iqbal et al, 2017). The use of a combination of GLCM feature vectors is able to distinguish US images of normal and diseased kidneys (Sharma and Virmani, 2017). Chen et al. (Chen et al., 2019) provided a comprehensive CKD staging analysis and classification method based on US image texture and achieved good classification performance. For changes in renal function after transplantation, US image texture had good correlation with sCr and was significantly different between subjects (rejected and non-rejected allografts), with good classification performance (Ardakani et al., 2017).

Texture analysis was also applied to kidney scintigraphy images. Ohata et al. extracted several textures and used different machine learning techniques to achieve the best accuracy for CKD stage classification (stages 1, 2 vs. stages 3-5) (Ohata et al., 2019). Ardakani et al. used textures to detect post-transplant kidney status and found that textures improved clinical diagnosis by providing good classification performance (Ardakani et al., 2018).

Applying texture analysis to clinical practice and research studies faces many challenges, such as the large impact of signal intensity and heterogeneity quantification, which requires standardization of image acquisition protocols and a priori normalization and intensity correction of images when needed (Timothy L. Timothy L. Kline et al., 2017). Of note, studies assessing the potential of texture analysis in the assessment of renal damage in CKD (beyond correlation with histology and pathology) remain limited and remain to be explored. Therefore, it is recommended to further verify its potential as a new dimension in CKD management. The methods proposed in Table 2 do not yet provide a complete and reliable method to aid clinical practice. These methods must be evaluated in large and multicenter studies that employ standardized imaging protocols with different modalities, apply standardized image intensity normalization patterns, respect patient preparation before image acquisition (e.g. fasting, check hydration status), and select the most reproducible and important textures that are more strongly related to kidney tissue histology and pathology than eGFR. Once these protocols are evaluated and proven to provide good performance in CKD detection and staging without misclassification, they will be able to be incorporated into clinical practice in terms of diagnosis and prognosis.

4.2 Deep learning

4.2.1 Principle

Deep learning is a part of machine learning that is inspired by the way the brain completes learning tasks and attempts to imitate biological neural networks (Goodfellow et al, 2016). It employs multi-layered artificial neural networks through mathematically interconnected nodes (Goodfellow et al., 2016). During the training process , the weights connecting these nodes are adjusted according to the optimization equation until the neural network learns well . Deep learning changes the traditional feature extraction process and subsequently uses traditional machine learning algorithms to transform it into a simple input-output process with a complex deep architecture that can achieve internal deep feature extraction (Kavur et al, 2020). Recent deep convolutional neural networks (ConvNets, cnn) are constructed with 10 to 20 layers of linear units, hundreds of millions of weights, and billions of connections between units (Lecun et al, 2015). The CNN architecture supports multiple array data processing including 1D signals, 2D images, audio and 3D videos (Lecun et al, 2015). The multi-layer composition of CNN can learn feature hierarchies without relying on hand-crafted features (Sharma et al., 2017a). In terms of image classification, CNN takes an image as input and passes the raw pixels through a convolutional filter to convert them into class scores (Sharma et al, 2017a).

Deep learning has been successfully used for organ segmentation (Roth et al., 2015; Zheng et al., 2017), total renal volume determination (Timothy L. Timothy L Kline et al., 2017; Sharma et al., 2017b), chronic Myocardial delineation (Zhang et al., 2019), cerebral microbleed detection (yd.; Zhang et al, 2018), and pulmonary nodule classification (Ciompi et al, 2015)

Deep networks were originally informed by medical clinical data, including attributes such as age, blood pressure, blood glucose, serum creatinine, and outperformed all traditional machine learning techniques (Kriplani et al., 2019; Saha et al., 2019; Shankar et al., 2018 ). Furthermore, CNN was found to be a powerful tool for glomerular localization in biopsy sections (Bukowy et al., 2018; Kannan et al., 2019; Marsh et al., 2018) and outperformed fibrosis based on pathology estimates in predicting fibrosis stage. Scored classifier (Kolachalama et al., 2018).

In terms of renal function assessment, CKD prediction, and diagnosis, deep learning or, more specifically, transfer learning, whether used alone or together with texture branching networks, shows good promise , as shown in Table 3. Recent research has focused on applying deep networks to different clinical applications, including differentiation of healthy kidneys from CKD kidneys, differentiation of normal kidneys from congenitally abnormal kidneys, classification of different kidney diseases such as kidney stones, cysts and tumors, and most interestingly Yes, eGFR is predicted.

4.2.2 Data availability

When training data is limited or difficult and expensive to collect, there is a need for neural networks that can overcome the problems caused by lack of data. Data augmentation is a necessary step before training a model. It increases the diversity of training samples and prevents the training model from overfitting (Sharma et al, 2017a). To expand the limited data, several methods were employed. Here we discuss the use of transfer learning and data augmentation by applying image transformations on existing images or by using an advanced topic in deep learning, deep generative adversarial networks (GANs).

Transfer learning aims to improve learning performance in the target domain by transferring information from related source domains (Weiss et al, 2016). For example, if a person has acquired good musical knowledge by playing the guitar and wants to learn to play the piano, by transferring his knowledge to the task of learning to play the piano, his Learning efficiency will be higher (Weiss et al, 2016). Therefore, transfer learning is possible.

Considering that deep networks require huge amounts of data during the training phase, data augmentation is widely used to enrich training data sets. Data augmentation is the artificial expansion of dataset samples by performing classic image transformations on existing samples, including random grayscale transformation (<3%) of pixels, cropping, rotation, scaling and movement (Hao et al., 2019; Pavinkurve et al., 2019), applying low-frequency intensity changes (Sharma et al., 2017a), or more complex image transformation algorithms such as radial transform sampling (H Salehinejad et al., 2018; Hojjat Salehinejad et al., 2018). Data can also be enhanced by applying deformable image registration (Yin et al, 2019).

Some studies have used transfer learning for imaging assessment of renal status. The researchers retrained a CNN that was previously trained on photographic images from the ImageNet challenge to classify ultrasound kidney images (Cheng and Malhi, 2017; Hao et al., 2019; Kuo et al., 2019 ; Zheng et al., 2019). The data were also augmented to obtain larger sample sizes required for deep networks. Cheng et al. demonstrated the effectiveness of transfer learning for classifying abdominal images for different diseases, including end-stage renal disease, liver and bladder diseases (Cheng and Malhi, 2017). Neural networks sometimes outperform radiologists in disease identification performance (Cheng and Malhi, 2017). In their recent study, Kuo et al. proposed a deep neural network model to estimate artificial intelligence-based GFR (AI-GFR) and detect CKD from kidney US images (Kuo et al., 2019). The authors extended their data with augmentation and employed a transfer learning approach with a pre-trained CNN model (ResNet), which has surpassed nephrologists' classification of CKD and non-CKD patients and improved between AI and scr-based eGFR. A good correlation was achieved between them (Kuo et al., 2019).

GAN is a part of deep generative modeling that takes input training samples from some distribution and learns a model that represents that distribution. gan relies on a generator and discriminator with multiple encoding and decoding layers. During GAN training, the generator attempts to create simulations (synthetic images) of the data, while the discriminator attempts to identify real data from the fake data created by the generator. Training continues until the discriminator cannot detect any differences between fake and real data . gan has attracted great interest in the field of medical imaging and has been used in different applications, including image synthesis (Lutnick et al., 2020; Sivanesan et al., 2019), image translation (Murali et al., 2020), image super resolution (Mahapatra and Bozorgtabar, 2019) and image conversion (Chandrashekar et al., 2019). For kidney images, GAN was used to generate real-like kidney micro-anatomical images (Murali et al, 2020). The authors have used cyclic gallium nitride to create artificial staining effects without the need to physically tamper with histopathology sections. Their proposed GAN was shown to be able to translate different types of renal pathology staining (such as converting hematoxylin and eosin staining into periodic acid Schiff staining). The team also used the concept of GANs to generate realistic synthetic images of kidney biopsies (Lutnick et al, 2020). Most interestingly, GANs were used by (Sivanesan et al., 2019) to create synthetic images of the United States as shown in Figure 7, thus extending their limited dataset.

4.2.3 Discussion on texture descriptors

Interestingly, the textures are combined with typical CNNs . New models are developed that combine deep networks and texture features as residual structures. This model uses a texture-branched CNN to extract depth features and texture features from the image, and uses the fused information for classification

Zheng et al. mixed transfer learning and traditional imaging features, including geometric features of oriented gradients (HOG) and histogram features (Zheng et al., 2019). The model they proposed has successfully distinguished between normal kidneys and kidneys with urinary tracts (Zheng et al., 2019). Likewise, to improve decision-making and classification performance, HAO et al. proposed adding texture features to CNN (ResNet), a method that provides additional descriptors (HAO et al., 2019). The authors used data augmentation before the transfer learning method and the residual texture branch to obtain a multi-level descriptor model that mixes depth features and domain texture features (Hao et al., 2019). The model they developed outperformed using texture features or CNN features alone with high accuracy and excellent sensitivity, and showed its ability as computer-aided CKD screening (Hao et al, 2019).

Artificial intelligence applies machine learning and deep learning algorithms to clinical research. Through research on AI-GFR, it was found that artificial intelligence is a useful non-invasive assessment tool, successfully transforming kidney imaging into a real-time screening tool and kidney function evaluator. . The combination of transfer learning and traditional image features provides a high-performance robust classifier. Further research into the application of deep learning to other imaging modalities is needed, and its effectiveness in detecting CKD and predicting its staging needs to be further verified.

4.3 Artificial Intelligence in Kidney Segmentation

4.3.1 Automatic kidney segmentation

Automatic kidney segmentation is a crucial task because manual delineation of kidney tissue is time-consuming and subject-dependent. Segmentation is a key step in abdominal image analysis, and its applications include surgical planning, computer-aided monitoring, qualitative or quantitative feature extraction, image-guided intervention, etc. (Conze et al, 2020). Segmentation of kidneys can serve radiologists before their qualitative assessment and is required before feature extraction from input images in order to obtain fully automated software tools for kidney detection and status assessment. Different segmentation methods have been used in clinical studies to avoid the effort of manual delineation of kidneys and observer intervention.

Kidney identification is generally performed semi-automatically or automatically . Since abdominal images have similar grayscale intensities and large changes in organ shape and location, semi-automatic segmentation is usually used. It requires multiple intervention mechanisms, such as identification of initial seeds, positioning of kidney tissue and samples within the background, pre-segmentation of kidneys with circular outlines, defining parameter ranges (Hammon et al., 2016; Hu et al., 2012; Mortensen and Barrett, 1998; Sandmair et al., 2016; Torres et al., 2018). However, these interventions are operator dependent, requiring additional analyzes to test segmentation reproducibility and reveal important inter- and intra-observer agreement and reliability. A limited number of experts are often available to perform such analyses, which prevents generalization of consistency and reproducibility (Kavur et al, 2020). Furthermore, user interaction takes time, which can be tedious in challenging tasks (Kavur et al, 2020). Therefore, many studies aim to convert user intervention into automatic by applying image processing and traditional machine learning algorithms that require manual feature extraction (Vasanthselvakumar et al., 2020), (Gao and bolang, 2010), (Akbari and Fei, 2013) Interaction. For more details on applying kidney semi-automatic segmentation strategies on MRI, US and CT imaging, see (Torres et al, 2018).

4.3.2 Traditional methods of automatic kidney segmentation

In terms of developing a fully automatic kidney segmentation framework, several traditional segmentation methods have been proposed in the past decade. For kidney segmentation in MRI images, threshold segmentation and shape detection, probabilistic shape models, Bayesian probability maps, unsupervised classification and deformable models were used. In (Will et al., 2014), the authors proposed efficient and simple segmentation of the renal cortex on T1 and T2 MR images using threshold segmentation and shape detection techniques. However, this method is limited by its dependence on MR image quality and visibility. (Shehata et al., 2018) used 3D probabilistic shape models on DWI images to evolve 3D geometrically deformable models that achieve high similarity coefficients with manual segmentation (DICE scoring or DSC). Despite satisfactory results, the developed model suffers from limitations of insufficient reliability and long running time. (Gloger et al., 2012) achieved a DSC of 0.90/0.89 for the left and right renal parenchyma on DCE images. The authors used fine probabilistic maps and combined outer cortical edge alignment to develop a 3D segmentation framework for fully automated renal parenchymal volume measurement, excluding parenchymal cysts, which can be used in clinical applications and epidemiological studies. However, the proposed method is specifically designed for epidemiological studies, requiring data and domain knowledge to be adapted during the training phase before extending the framework for kidney volume determination. Furthermore, clustering methods were used on DCE images to identify multiple kidney structures (such as cortex, medulla, and pelvis) (Chevaillier et al., 2011; Li et al., 2012; Yang et al., 2016; x 2015; Zöllner et al. . , 2011). Indeed, the identification of renal structures can be aided by information on the temporal behavior of the contrast agent and on changes in pixel intensity over time. In their method, pixels are classified according to the intensity evolution of clusters, including K-means (Chevaillier et al., 2011; Yang et al., 2016; x 2015; Zöllner et al., 2011), growing neural gases (Chevaillier et al., 2011 ), wavelet-based (Li et al., 2012; Zöllner et al., 2011) and Gaussian mixture clustering (Zöllner et al., 2011). However, clustering is based on the behavior of pixels under the contrast agent, which is not necessarily used in other imaging sequences or modalities. Additionally, active contour (AC) or deformable models use energy constraints and forces in the image to isolate regions of interest (Al-Shamasneh et al., 2020; LL Li et al., 2014). The AC model is based on partial differential equations and variational models. The main idea of ​​AC in image segmentation is to start from an arbitrary boundary (closed curve), then iteratively update the curve through shrinkage, and move the curve through image-driven power to accurately detect object boundaries within the image (Hoang Ngan Le et al, 2020) . Li et al. successfully segmented the complex kidney contours on DCE MRI images by utilizing the geometric active contours provided by a multi-scale edge detection algorithm that segmented non-uniform regions (LL Li et al, 2014). Recently, Al-Shamasneh et al. proposed a new active contour model for kidney segmentation on low-contrast MR images (Al-Shamasneh et al, 2020). Their model uses a new fractional function (mittagg-leffler function) to achieve energy minimization, which is better than other methods, such as the Chan-Vese active contour model (CV model), which was proposed by Ibrahim et al., using Lai Particular fractional functions and deep serial network segmentation (Chan and Vese, 2001; Ibrahim et al, 2018). The results show that this method has high segmentation accuracy (98.95%) and DSC (0.93). , 2014). The AC model is based on partial differential equations and variational models. The main idea of ​​AC in image segmentation is to start from an arbitrary boundary (closed curve), then iteratively update the curve through shrinkage, and move the curve through image-driven power to accurately detect object boundaries within the image (Hoang Ngan Le et al, 2020) . Li et al. successfully segmented the complex kidney contours on DCE MRI images by utilizing the geometric active contours provided by a multi-scale edge detection algorithm that segmented non-uniform regions (LL Li et al, 2014). Recently, Al-Shamasneh et al. proposed a new active contour model for kidney segmentation on low-contrast MR images (Al-Shamasneh et al, 2020). Their model uses a new fractional function (mittagg-leffler function) to achieve energy minimization, which is better than other methods, such as the Chan-Vese active contour model (CV model), which was proposed by Ibrahim et al., using Lai Particular fractional functions and deep serial network segmentation (Chan and Vese, 2001; Ibrahim et al, 2018). The results show that this method has high segmentation accuracy (98.95%) and DSC (0.93). , 2014). The AC model is based on partial differential equations and variational models. The main idea of ​​AC in image segmentation is to start from an arbitrary boundary (closed curve), then iteratively update the curve through shrinkage, and move the curve through image-driven power to accurately detect object boundaries within the image (Hoang Ngan Le et al, 2020) . Li et al. successfully segmented the complex kidney contours on DCE MRI images by utilizing the geometric active contours provided by a multi-scale edge detection algorithm that segmented non-uniform regions (LL Li et al, 2014). Recently, Al-Shamasneh et al. proposed a new active contour model for kidney segmentation on low-contrast MR images (Al-Shamasneh et al, 2020). Their model uses a new fractional function (mittagg-leffler function) to achieve energy minimization, which is better than other methods, such as the Chan-Vese active contour model (CV model), which was proposed by Ibrahim et al., using Lai Particular fractional functions and deep serial network segmentation (Chan and Vese, 2001; Ibrahim et al, 2018). The results show that this method has high segmentation accuracy (98.95%) and DSC (0.93). 2018). The results show that this method has high segmentation accuracy (98.95%) and DSC (0.93). 2018). The results show that this method has high segmentation accuracy (98.95%) and DSC (0.93).

In addition, shape detection technology and deformable models were used to perform kidney segmentation on US images. In (Marsousi et al., 2017), kidneys were detected by fitting a 3D shape kidney model on a 3D US volume. Then, the fitted model evolution level set function was used to delineate the kidney boundaries. The accuracy of the author's method is 97.48% and the DSC is 0.81, which is better than other methods (Ardon et al, 2015; Marsousi et al, 2014). This method fails to detect kidney volume in low-quality ultrasound images. (Yang et al., 2012) adopted a distance regularized level set deformation model followed by a shape prior for smooth boundary detection. Their model showed 95% sensitivity and 95% specificity. In (Huang et al., 2013), a new active contour framework is proposed, in which fast segmentation with an error of 0.028 is achieved by performing a convex relaxation of the energy function. Although supervised classification has proven to be a successful segmentation method, it is still limited by manual feature extraction. Although the deformation model is one of the most widely used medical image segmentation methods in recent decades, it has some limitations. AC does not need to learn attributes from training images. Therefore, it has difficulty handling occlusions and noise. Furthermore, it represents an unsupervised framework that lacks a way to process labeled images in a supervised manner. Therefore, it gives unpredictable segmentation results. Finally, the deformable model strongly depends on several parameters selected from experimental results (Hoang Ngan Le et al, 2020).

4.3.3 Deep networks for automatic kidney segmentation

On the other hand, recent studies have shown that the system structure provided by deep networks is able to turn segmentation into a fully automatic process with high accuracy and repeatability, requiring neither intervention nor manual features. In addition to the success of CNNs in kidney failure detection and patient classification, whole kidney segmentation is a fundamental problem that has been addressed recently, as shown in Table 4. Most recent methods are based on semantic segmentation using fully convolutional networks (FCN) or convolutional encoder-decoder networks (CED). The encoding and decoding parts of CED allow distinguishing an image pixel because it does or does not belong to a specified region of interest (pixel wise segmentation). Figure 8 (left panel) gives an example of a CED architecture.

Kidney segmentation using deep learning was tested on MR, US and CT images. Bevilacqua et al., in their study, achieved an overall segmentation accuracy of 86% on MR images of kidneys with ADPKD (Bevilacqua et al., 2018). The authors tested different CNN-based segmentation methods on their augmented dataset. Full-image segmentation based on FCN has surpassed CED (SegNet with VGG-16 as the encoder) segmentation, and the introduction of region-based CNN (R-CNN) before semantic segmentation did not improve the rendering performance. In addition, Yin et al. developed a novel boundary distance deep network based on transfer learning for segmenting US kidney images of children with congenital kidney and urinary tract anomalies (CAKUT) and controls, with an accuracy of 98.9% (Yin et al. , 2019). The CNN (VGG-16) pre-trained on ImageNet is used to extract deep features, and the boundary map is learned through the boundary distance regression network. The predicted maps are then classified as kidney pixels or non-kidney pixels using a pixel classification network. Furthermore, Sharma et al. used a CNN following the VGG-16 architecture to generate scoring maps based on pixel classification (see Figure 8) and achieved a DICE score of 0.86 between proposed and manual delineation of ADPKD kidneys on CT images (Sharma et al. , 2017a). Similarly, Thong et al. proposed a fully automatic framework for patch-based contrast-enhanced kidney segmentation of CT scans trained on CNN. The proposed model achieved a similarity score between manual and automatic kidney delineation, with similarity scores exceeding 0.9 for both left and right kidneys (Thong et al, 2016).

Most interestingly, the Comprehensive Healthy Abdominal Organ Segmentation (CHAOS) benchmark aims to develop algorithms capable of segmenting multiple organs including kidneys, liver, and spleen on CT and MRI images (T1-DUAL and T2-SPIR sequences) from public datasets. Deep network of abdominal organs and provides a realistic organ delineation (Conze et al., n.d.). The CHAOS challenge involves different tasks, ranging from liver segmentation on CT or MRI modalities only, to the most complex task of abdominal organ segmentation on CT and MRI modalities using a single network architecture (pretrained or untrained) (Conze et al, 2020). All but one of the participating teams used extensions of the U-Net deep network and all achieved high segmentation processes. A recent cross-modal multi-organ segmentation study for CHAOS showed that CHAOS performed well on multi-organ segmentation tasks and won first place in the three competition categories of liver MR, liver CT and multi-organ MR segmentation (see figure 9) (Conze et al., 2020). The authors tested different pipelines based on different pre-trained networks. The highest left and right kidney segmentation scores were obtained when U-Net was pre-trained using cascades associated with conditional generative adversarial networks (cGAN). This structure enhances the network's segmentation ability for multiple organs and has good generalization ability. In short, in conditional GAN, the generator learns to create synthetic images under certain specific conditions. In terms of segmentation tasks, the generator, as the name suggests, will generate masks through its encoding and decoding layers, while the discriminator evaluates whether the generated masks are real or not. Therefore, the adversarial network will be able to distinguish between real and synthetic depictions and force the generator to create masks that are as realistic as possible. Cascading ced is used to take advantage of multi-level contextual information (Conze et al, 2020).

Automatic kidney segmentation based on deep networks has been reported to be promising . Deep networks take advantage of supervised methods that do not require hand-crafted features. Intervention is limited to setting network hyperparameters (batch size, number of iterations) that may affect model convergence. Most interestingly, such networks can provide appropriate solutions to cross-modal problems. Another approach is to train a single model to segment MRI and CT images.

There are several limitations to using deep learning for kidney segmentation that need to be mentioned. Deep networks are supervised methods that require large amounts of real data for training, and synthetic data can be generated by using various data augmentation modes and GANs. Furthermore, transfer learning provides a way to accelerate model convergence during the training phase (Conze et al, 2020). Furthermore, it is notoriously difficult to determine an appropriate architecture for a specific semantic segmentation problem. Therefore, testing several network architectures and comparing segmentation results is the best way to determine which model provides the highest segmentation performance. What should be considered is the need for large memory and powerful graphics cards for deep networks, as well as the empirical need to find optimal convergence parameters (Kavur et al, 2020).

Although many kidney segmentation methods and deep models have been developed, research in this area is still limited and more models need to be introduced in future work. Deep segmentation frameworks such as DeepMedic (Kamnitsas et al., 2017) and NiftyNet (Gibson et al., 2018) must be explored. Networks such as VNet (Milletari et al., 2016), ScaleNet (Fidon et al., 2017), and HighRes3dNet (W. Li et al., 2017) must be evaluated for kidney segmentation. Interestingly, adding active contour models to deep network frameworks can inherit the advantages of both. Several methods have been proposed, (i) AC can be used as a post-processing tool after deep networks, (ii) in deep networks, it can replace fully connected layers for segmentation, (iii) energy minimization of AC can be used as Network loss function, (vi) Deep networks can learn parameters from AC. These methods have not yet been explored in the field of kidney segmentation (Hoang Ngan Le et al, 2020). Best of all, deep networks can work in parallel given the necessary computing power. Therefore, their results can be combined through an integrated system (Kavur et al, 2020) to obtain better performance. The Ensemble of Multiple Models and Architectures (EMMA) model is a good example of such a fusion system (Kamnitsas et al., 2018). The application of this promising approach to kidney segmentation has yet to be described. Finally, deep learning-based segmentation of renal structures (e.g., medulla, cortex, pelvis) remains to be explored.

5 Conclusion

Renal disease is characterized by alterations in kidney macrostructure (including kidney volume and CMD) and microstructure (including fibrotic component and lipid fraction). Developing methods for early diagnosis and prediction of CKD remains a challenging problem that could reduce the cost of disease treatment and slow the progression of kidney damage.

MRI has proven to be a powerful tool to evaluate renal tissue by assessing renal function and structure of both kidneys. MRI is embedded with sequences that reflect different kidney properties and functions, including diffusion, perfusion, oxygenation, tissue elastography, hemodynamics, and more. Although MRI sequences provide the opportunity to assess the microvascular and microstructural integrity of both kidneys, there are still limitations. Other factors may have an impact on the measured markers (e.g., urinary flow rate, edema and DWI medications, BOLD intravascular volume and canalicular function). disorders, ASL intravascular volume and medications, etc.). In addition, human studies evaluating the effectiveness of MT and T1 and T2 localization are needed. Multiparametric MRI studies can better reflect the correlation between all renal function measurements and allow the selection of the most effective MRI sequence to detect renal failure. Taking advantage of its low cost and availability, UE also monitors CKD by measuring kidney stiffness like MRE, but the conflicting results obtained in UE studies make MRE more suitable for kidney stiffness assessment. Although CT and scintigraphy are effective in testing renal function, they are still limited to radiation exposure and contrast agent injection.

In terms of routine renal function tests, in order to prevent the kidneys from being affected, US as well as UE and CT provide suitable imaging techniques that can replace biopsy and GFR examination. These modalities provide low-cost, non-invasive, clinically available and short-time examination techniques. These imaging techniques can provide rapid assessment tools when strong suspicion of CKD arises. What people should be aware of is the radiation exposure caused by CT. Given that changes in renal tissue occur microscopically and often require biopsy for evaluation, nephrology should take advantage of functional MRI in high-risk patients. MRI can predict renal failure associated with dysfunction because it safely gives functional and structural parameters. In patients with CKD or high-risk patients, contrast-enhanced and radiation exposure imaging should be eliminated due to its side effects, especially when the kidneys are already affected.

Texture analysis has the advantage of reducing inter-observer variability because the analysis covers the entire renal parenchyma, which is a spatially heterogeneous tissue due to its internal structure. Texture expands the possibilities of medical imaging and can serve as complementary evidence to traditional qualitative markers. Artificial intelligence detects kidney deformation earlier. It uses machine learning techniques and deep neural networks, which perform well in early CKD detection and GFR estimation through AI-GFR. Artificial intelligence has demonstrated its ability to transform traditional medical imaging into real-time screening tools that can help doctors make clinical decisions. The texture branch provides ResNet network performance that outperforms other networks and is a promising architecture that must be tested on other imaging modalities. While the ability of deep networks to possess fully automated kidney segmentation and disease classification strategies is associated with high performance, CNNs lack the large amounts of data overcome using transfer learning methods and data augmentation for training, and require high computing power to implement (e.g. larger memory, powerful graphics card). In addition, deep networks require strong experience to select optimal convergence parameters and set optimization parameters.

Even though they achieved encouraging results in early detection of CKD and GFR estimation, the proposed models based on texture analysis and deep networks are still insufficient to be incorporated into clinical practice until excellent detection and estimation performance are achieved. An accuracy rate of <99% means that there are still misclassifications of some cases. Future research efforts should explore successful literature models in large follow-up and multicenter studies that use standardized imaging protocols and homogeneous patient preparation scenarios (e.g. fasting, standardized hydration status, verification of salt and medication intake prior to imaging) , and balance the distribution of positive and negative samples. The most reproducible and distinctive features need to be selected. The association between texture and deep features must be evaluated to gain a deep understanding of their role in the detection and staging of renal damage.

Researchers must strive to perform automated kidney segmentation using different imaging modalities before disease diagnosis to enrich fully automated renal dysfunction detection networks with high accuracy and efficiency. VGG-16 and cGAN/U-Net pre-trained on the ImageNet challenge show promise in segmentation tasks. The use of a single framework to achieve multi-modal segmentation and CKD assessment must be addressed because cross-modal training is still more challenging than individual learning. Future work must include evaluation of different deep networks (DeepMedic, ScaleNet, VNet and HighRes3dNet) for cross-modal kidney segmentation. In addition, the application of deep networks combined with active contours in the field of kidney segmentation also needs to be studied. Furthermore, the effectiveness of integrated fusion systems in this context remains to be tested. To incorporate the kidney segmentation depth model into clinical applications, it must have good depiction performance (such as DICE score, accuracy, sensitivity).

CKD staging challenges must be further evaluated through deep neural networks. Integrating texture from other imaging modalities into these networks requires further validation to test their effectiveness in clinical decision-making and outcome prediction.

Finally, in terms of clinical workflow, the gap between research and real-life tools needs to be filled. The successful segmentation and kidney disease assessment solutions proposed by the researchers must be implemented in real-world applications . Such a framework should be easy for clinicians to use in their daily clinical work.