introduce
Medical image segmentation is a critical process in the healthcare industry. It involves dividing a digital medical image into parts to simplify and/or change the representation of the image into something more meaningful and easier to analyze. Over the years, various technologies have been developed to enhance this process, and one of the latest advancements in this field is MedLSAM.
Keywords: SAM·Medical image segmentation·Contrast learning
Learn about MedLSAM
MedLSAM is a novel model developed by researchers to simplify the annotation process for medical image segmentation. This model introduces a unique few-shot localization framework that can identify any target anatomical part in the body. This is achieved by using only six extreme points in three directions on several templates. The result is a more streamlined and efficient process that significantly reduces the time and effort required for medical image annotation. The few-shot localization framework in MedLSAM, also known as MedLAM, maintains strong similarity in observations based on the spatial distribution of organs across different individuals. This observation enables MedLAM to locate any target anatomical site in the body.
Segmentation Arbitrary Model (SAM) has recently become a breakthrough model in the field of image segmentation. However, both the original SAM and its medical adaptation require slice-by-slice annotation, which directly increases the annotation workload with the size of the dataset. The authors proposed MedLSAM to solve this problem and ensure a constant annotation workload regardless of the dataset size, thus simplifying the annotation process. Our model introduces a several-shot localization framework capable of localizing any target anatomical part in the body. To achieve this goal, the authors developed a Localized Any Model (MedLAM) for 3D medical images, leveraging two self-supervised tasks: relative distance regression (RDR) and multi-scale across a comprehensive dataset of 14,012 CT scans. Similarity (MSS). Then, the authors established an accurate segmentation method by integrating MedLAM with SAM. By annotating only six extreme points in three directions on several templates, our model can autonomously identify target anatomical regions on all planned annotated data. This enables our framework to generate a 2D bounding box for each slice of the image