AI in Radiology: Multimodal Imaging, MedSAM, Vision Transformers and Clinical XAI

To develop practical skills in building, evaluating and explaining AI models for multimodal radiology and medical image analysis

Participants will learn to:

• Process DICOM and NIfTI medical images.
• Normalize, align and resample CT and MRI scans.
• Build multimodal medical imaging pipelines.
• Apply MedSAM and Swin UNETR for segmentation.
• Perform multi-label chest X-ray classification.
• Generate Grad-CAM and Integrated Gradients explanations.
• Evaluate models using Dice Score, AUROC, precision, recall and F1-score.
• Prepare publication-ready figures and validation reports.

🗓️ Day 1: Multi-Modal Radiology Data Pipeline Engineering

• Objective: Build standardized and memory-efficient pipelines for processing multi-modal medical imaging data.
• Introduction to CT, MRI, X-ray, and clinical data fusion.
• Understanding DICOM and NIfTI medical imaging formats.
• Managing LPS/RAS orientation and spatial alignment.
• Applying intensity normalization and Hounsfield Unit windowing.
• Comparing early-fusion and late-fusion architectures.
• Isotropic voxel resampling for consistent 3D analysis.
• Patch sampling and sliding-window inference for large volumes.

🛠️ Hands-on Lab (Google Colab)

• Task: Build a MONAI pipeline to load, normalize, resample, and spatially align paired T1/T2 MRI and CT scans.
• Tools Covered: MONAI, SimpleITK, PyTorch, TorchIO, and Google Colab.

🗓️ Day 2: Vision Transformers, MedSAM and Automated Pathology Detection

• Objective: Apply transformer architectures and medical foundation models for segmentation and multi-label diagnosis.
• Evolution from U-Net to transformer-based medical image segmentation.
• Architecture and clinical applications of Swin UNETR.
• Introduction to MedSAM and medical foundation models.
• Using bounding-box, point, and mask-based segmentation prompts.
• Zero-shot segmentation of lesions and anatomical regions.
• Transfer learning with pre-trained radiology models.
• Multi-label detection of pneumonia, pneumothorax, and pulmonary nodules.

🛠️ Hands-on Lab (Google Colab)

• Task: Segment lung nodules using MedSAM prompts and perform multi-label chest X-ray classification using TorchXRayVision.
• Tools Covered: MedSAM, TorchXRayVision, Swin UNETR, MONAI, Hugging Face, and Google Colab.

🗓️ Day 3: Explainable AI, Clinical Validation and Research Publication

• Objective: Interpret medical AI predictions, evaluate model performance, and prepare reproducible research outputs.
• Understanding the importance of explainability in clinical AI systems.
• Creating Grad-CAM and Layer-CAM visualizations.
• Applying Integrated Gradients to multi-modal medical imaging models.
• Evaluating segmentation using Dice Score and Hausdorff Distance.
• Evaluating classification using AUROC, precision, recall, and F1-score.
• Designing ablation studies and comparative experiments.
• Creating publication-ready figures and reproducible research workflows.

🛠️ Hands-on Lab (Google Colab)

• Task: Generate saliency maps using Captum and Grad-CAM, create 300-DPI radiology visualizations, and produce statistical validation reports.
• Tools Covered: Captum, PyTorch Grad-CAM, scikit-learn, MONAI Label, Matplotlib, Seaborn, and Google Colab.

• Radiologists and clinical researchers
• PhD scholars and postgraduate students
• Academicians and faculty members
• Biomedical and medical imaging researchers
• Healthcare data scientists
• AI and machine learning professionals
• Biomedical engineers
• Medical imaging and diagnostic-AI professionals

By the end of the workshop, participants will be able to:

• Build standardized MONAI imaging pipelines.
• Segment lesions and anatomical regions using MedSAM.
• Apply Vision Transformers to medical imaging tasks.
• Classify radiological abnormalities from chest X-rays.
• Interpret model predictions using clinical XAI methods.
• Evaluate segmentation and classification performance.
• Create reproducible research workflows and publication-ready outputs.