Explainable AI (XAI) & Clinical PlausibilityActive Benchmark / Project 6
License: MITExplainability & Clinical Plausibility of Deep Learning on Low-Field African Brain MRI
Evaluating Anatomical Attention vs. Shortcut Learning across Low-Field Scanners with Neuroradiology Expert Auditing
Python 3.11+PyTorchMONAIGrad-CAM++Integrated GradientsOcclusion SensitivityKernelSHAPResNet-18DenseNet-121ConvNeXt Tiny
Core Research Ecosystem Vision
Evaluating whether deep learning models trained on heterogeneous low-field African clinical MRI base decisions on medically plausible neuroanatomy or spurious imaging shortcuts using a 5-point neuroradiologist rating rubric and quantitative anatomical IoU.
5-Attribution Explainability Framework
Gradient-based
Grad-CAM
Coarse class activation mapping from final convolutional layer gradients.
Gradient-based
Grad-CAM++
Enhanced localization handling multiple object instances of identical classes.
Pixel-Attribution
Integrated Gradients
Path-integrated gradients relative to a baseline image.
Perturbation-based
Occlusion Sensitivity
Systematic patch masking to measure prediction probability changes.
Game-Theoretic
SHAP (Kernel/Deep)
Shapley values assigning exact feature importance weights.
Neuroradiologist 5-Point Likert RubricClinical Plausibility Criteria
Score 1 (Entirely Implausible)Heatmap focuses purely on background, skull, or imaging artifacts.
Score 2 (Mostly Implausible)Attention is mostly on non-diagnostic regions with minimal target overlap.
Score 3 (Indeterminate)Focus is diffuse across normal brain structures showing no diagnostic preference.
Score 4 (Mostly Plausible)Major attention resides on expected diagnostic neuroanatomy with minor distraction.
Score 5 (Highly Plausible)Attention maps are strictly confined to pathological structures per diagnostic criteria.
Disease PathologyExpected ROI vs. Shortcut Trap
Hydrocephalus
Target ROI: Lateral ventricles, third ventricle, periventricular zones
Shortcut Trap: Skull edges, orbits, neck, background noise
Dementia
Target ROI: Hippocampal formation, temporal lobes, cortical sulci
Shortcut Trap: Occipital lobe, skull, cerebellomedullary cistern
Parkinson’s Disease
Target ROI: Midbrain region, substantia nigra locus
Shortcut Trap: Frontal cortex, orbits, scanner artifacts
Epilepsy
Target ROI: Hippocampus, mesial temporal lobes, cortical gyri
Shortcut Trap: Ventricles, skull, skull base
Healthy Controls
Target ROI: Distributed, non-focal normal anatomical structures
Shortcut Trap: Single-voxel spikes, background noise
Execution Commands & Pipeline Workflow
# 1. Train MONAI 3D DenseNet-121 Benchmark:
python src/models/cnn.py --model densenet121 --task task4 --epochs 50 --batch_size 4 --out_dir ./checkpoints/densenet
# 2. Evaluate Degradation Robustness & RRI Metrics:
python src/evaluation/metrics.py --evaluate_robustness --model_dir ./checkpoints/ --out_dir ./results/
# 3. Generate Grad-CAM Explainability Heatmaps under Perturbation:
python src/evaluation/explain.py --model_path ./checkpoints/densenet/best.pth --method gradcam --out_dir ./results/explanations/