Decode the Brain: Predict Cognitive Decline with AI

The overall aim of this workshop is to bridge the gap between neuroimaging data science and advanced machine learning. It equips participants with the theoretical frameworks and hands-on skills required to process raw, multimodal brain data (MRI and fMRI) and build, interpret, and deploy clinically valid deep learning models capable of predicting cognitive decline and pattern recognition.

• Establish Clinical & Neuroanatomical Foundations: Contextualize the clinical challenge of neurodegenerative disease staging (CN rightarrow MCI rightarrow AD) and isolate key affected brain regions like the hippocampus.

• Master Preprocessing and Harmonization Protocols: Train participants on medical image standards (DICOM, NIfTI, BIDS) and multi-site dataset normalization using ComBat harmonization to resolve scanner variability.

• Implement Classical & Deep Learning Architectures: Guide participants from feature-extraction-based classical ML models (SVM, Random Forest) up to advanced 3D Convolutional Neural Networks (3D ResNet) and Swin Transformers designed for volumetric brain scans.

• Enforce Model Interpretability: Provide rigorous training on medical AI interpretability tools (SHAP, GradCAM) to ensure machine learning predictions align closely with biological ground truths.

• Bridge the Gap to Clinical Deployment: Provide structural frameworks for clinical validation metrics, TRIPOD-AI reporting standards, FDA SaMD regulations, and building API inference skeletons.

DAY 1 — Brain Imaging Foundations + First Analysis 

• The clinical problem: cognitive decline pipeline (CN → MCI → AD)

• What MRI/fMRI measure + key regions (hippocampus, DMN, PCC)

• Data labelling and disease staging

• DICOM vs NIfTI vs BIDS — quick primer

• 6-step structural MRI pipeline + fMRI-specific steps

• ComBat harmonization (mention, don't deep-dive)

• OASIS dataset via nilearn + inspect NIfTI in Environment setup (Colab), Visualization of 3 planes (axial/coronal/sagittal), Extract hippocampal ROI (AAL atlas) and Hippocampal atrophy scatter plot + ANOVA across groups

DAY 2 — ML on Brain Imaging 

• Classical ML on features (RF, SVM) vs CNN vs 3D CNN vs GNNs

• Class imbalance solutions

• Clinical metrics (AUC, sensitivity, specificity)

• VBM + functional connectivity basics

• SHAP for brain region importance

• Training Random Forest + SVM (stratified CV), ROC curves + confusion matrices, Functional connectivity matrix from fMRI + vectorize

DAY 3 — Deep Learning + Deployment 

• 3D ResNet: architecture overview

• MONAI ecosystem (why, not deep setup)

• Transfer learning for small datasets

• Late fusion (CNN + clinical variables)

• Site-leakage, temporal-leakage

• Full clinical metrics: AUC, sensitivity, PPV, calibration

• FDA SaMD + TRIPOD-AI reporting (overview)

• Build 3D ResNet-10 in MONAI  and data pipeline (spacing, intensity, augmentation), Training loop (mixed precision + scheduler) and GradCAM on 3D CNN: saliency maps, Multimodal fusion: CNN + clinical variables and GradCAM on 3D CNN: saliency maps

• Set up a medical imaging AI environment using MONAI, PyTorch, Nibabel, and Nilearn.
• Build end-to-end MRI/fMRI data pipelines, including image loading, preprocessing, normalization, and augmentation.
• Work with neuroimaging data formats such as DICOM, NIfTI, and BIDS.
• Extract hippocampal ROI features, functional connectivity matrices, and VBM-based imaging features.
• Build and compare machine learning models such as Random Forest, SVM, and deep 3D CNNs for cognitive decline prediction.
• Evaluate models using clinical metrics such as AUC, sensitivity, specificity, PPV, ROC curves, and confusion matrices.
• Interpret AI predictions using SHAP and 3D GradCAM to identify important brain regions.
• Understand clinical AI deployment concepts including site leakage, bias, TRIPOD-AI reporting, FastAPI inference, and structured prediction reports.

• Undergraduate and Post Graduate Students
• Phd Scholars and Researchers
• Clinicians and Physicians
• Professors and Industrial Professionals
• Data scientists, AI/ML learners

Take Home Outcomes:

• QC report: Inspect 5 subjects across CN/MCI/AD groups for visible atrophy
• Pipeline run: Successfully preprocess 1 subject through fMRIPrep
• BIDS tree: Reorganize 10 subjects into valid BIDS structure using heudiconv
• Classifier trained: SVM + XGBoost on ADNI hippocampal + FC features
• SHAP analysis: Identify top 10 most predictive connectivity edges for MCI
• ROC curves: Compare SVM vs RF vs XGBoost AUC on held-out test set
• Trained 3D CNN: AD vs CN AUC target: >0.85 on OASIS-3 held-out set
• GradCAM map: Anatomically validate: activation peak in hippocampus
• Longitudinal: Extend to 2-timepoint OASIS-3 to predict conversion risk