Data preparation for artificial intelligence in medical imaging: A comprehensive guide to open-access platforms and tools
Data Preparation Process
(i) image acquisition at clinical sites, (ii) image de-identification to remove personal information and protect patient privacy, (iii) data management to control the quality of image and non-image information, (iv) image storage and management, and finally (v ) image annotation.
FAIR Data Guiding Principles: Findable, Accessible, Interoperable and Reusable
Data Classification: Any Protected Health Information (PHI) and Personally Identifiable Information (PII) linked to patient health information PII: Can include data from Electronic Health Records (EHRs), medical images, clinical and biological data, and health services Any other data collected by the provider
Data format: DICOM. Often referred to as metadata, images contain a header that contains information about the image sequence, hospital, provider, clinician, or patient information, among other things.
De-identification tools: The first three of FreeSurfer's mri_deface, pydeface, mridefacer and Quickshear are python libraries
List of requirements for data de-identification tools: remove unused identification information and keep it depending on the situation; train at clinical sites to avoid secondary transmission of data
ISO 25237: Adoption Standard for Existing Privacy Protections
detailed list
Content management tool: can perform format conversion and data test training set management
DICOM format such as DCMTK
GUI: Posda, DVTk, dcm4che, BrainVoyager, LONI Debabeler, the GUI of dcm2niix is MRIcroGL Posda, DVTk and dcm4che provide web\GUI interface
image storage
PACS: Store multimodal medical images (MRI, computed tomography CT, positron emission tomography PET, etc.) Easy access from multiple devices and locations. (Multiple backups Quality Assurance (QA) Workstation: Validates patient demographics and any other important attributes of the study. Archive (central storage device): Stores validated images along with any reports, measurements and other relevant information. Reading Workstation: Where radiologists review data and formulate a diagnosis.
Medical automation system interface: Hospital Information System (HIS), Electronic Medical Record (EMR) and Radiology Information System (RIS)
XNAT: Java web, search in Postgres database, XML-based data model, it can support any type of tabular data
Dicoogle: open source PACS archive, no database
Orthanc
Annotation of 3D images: 3D Slicer
Image Annotation Tool: 3D Slicer
ITK-SNAP: Labeling can be performed on all three orthogonal slices (axial, coronal and sagittal) and displayed as a 3D rendering
MITK: Ability to assist annotation servers and perform automatic segmentation
Horos Viewer: based on OsiriXTM and other open source medical image libraries
ImageJ: java, supports DICOM
Plugins:
Bio-Formats: Conversion of proprietary microscopy image data and metadata into standardized open formats
SciView: 3D Visualization and Virtual Reality Capabilities for Images and Meshes
MaMuT: annotations for browsing, annotating and managing large image data
Trainable Weka Segmentation: Produce pixel-based segmentation
crowd Cure and CMRAD Platform: Cloud
Medical Image Repository
Various sources of open access medical imaging databases by target organ and disease
TCIA: Multiorgan Imaging for Cancer Imaging. DICOM format, using National Biomedical Imaging Archive (NBIA) software ( https://imaging.nci.nih.gov ) as backbone
UK Biobank: clinical data such as electronic medical records, which has an imaging collection of more than 100,000 participants, including scans of the brain, heart, abdomen, bones and carotid arteries
AI development direction: (i) data enhancement and synthesis, (ii) Federated learning, (iii) reasonable use of AI, (iv) uncertainty estimation
Data augmentation and synthesis: use feasible basic strategies of geometric transformation, flipping, color modification, cropping, rotation, noise injection, and random erasure, as well as other more advanced techniques, including creating new synthetic images, such as generative adversarial networks
Federated learning: Algorithms are trained on multiple decentralized clinical sites holding local data samples without exchanging data samples. The locally trained AI results are then combined in a centralized location.
Fair Use: Routine clinical data collected at clinical sites may be flawed, biased (eg, gender imbalance), or prone to noise (eg, in the presence of image artifacts). Some progress can be achieved by defining algorithms that efficiently track these problems
Uncertainty: model accuracy, which indicates how dependent clinicians are on the software
