Neuroimaging & Medical AIActive Benchmark / Phase 3B

License: MIT

Beyond Research-Grade MRI: AI Robustness on Nigerian Clinical Scans

A Methodological Study of Model Calibration, Robustness, and Explainability under Real-World MRI Quality Constraints

Python 3.11+PyTorchMONAISimpleITKPyRadiomicsScikit-LearnResNet-18DenseNet-121

Core Research Ecosystem Vision

Systematically characterizing AI model robustness, calibration error, and explainability maps across progressive real-world MRI degradations using a Nigerian clinical brain dataset.

Dataset Specifications & Modality Coverage

Cohort

88 subjects (33 Control [37.5%], 33 Dementia [37.5%], 22 Parkinson’s [25.0%])

Scans

787 structural images (T1w, T2w, FLAIR) acquired using 1.5T and sub-Tesla clinical scanners

Scanner Environment

Variable modality coverage, multiple orientations (Axial, Coronal, Sagittal), real-world clinical quality variability

Sequence / CombinationAvailable Subjects (%)

T1w88 subjects (100.0%)

T2w82 subjects (93.2%)

FLAIR44 subjects (50.0%)

T1w + T2w82 subjects (93.2%)

T1w + FLAIR44 subjects (50.0%)

T2w + FLAIR44 subjects (50.0%)

T1w + T2w + FLAIR44 subjects (50.0%)

Multi-Task Clinical Benchmarks

Task 1: Control vs. Dementia

Separation of normal aging from cognitive decline.

Task 2: Control vs. Parkinson’s

Sensitivity to subcortical Parkinsonian structural alterations.

Task 3: Dementia vs. Parkinson’s

Differential diagnostics between distinct neurodegenerative pathways.

Task 4 (Primary): 3-Class Classification

Simultaneous multi-class categorisation (Control vs. Dementia vs. Parkinson’s).

Controlled Physical Degradation Axes (L0 → L3)

Gaussian Blur

Simulates spatial resolution limits and hardware reconstruction smoothing.

Rician Noise

Signal-dependent magnitude noise modeling real-world MRI magnitude physics.

Slice Downsampling

Anisotropic resolution reduction (z-spacing 2.0mm–5.0mm) simulating thick clinical slices.

Motion Ringing

k-space phase perturbations generating realistic subject motion ghosting.

Phase 3A Radiomics Robustness Benchmark Results

Corruption Tier

Macro F1

RRI

Impact

Clean Baseline (L0)

0.5518

1.0000

Clean reference baseline

Gaussian Blur L1

0.4019

0.7283

Mild decay (-27.2%)

Gaussian Blur L2

0.2394

0.4338

Steep decay (-56.6%)

Gaussian Blur L3

0.2139

0.3876

Floor performance (-61.2%)

Motion Ringing L1-L3

0.1898

0.3439

Catastrophic performance drop (-65.6%)

Rician Noise L1

0.4005

0.7257

Moderate decay (-27.4%)

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/

FILIMA PATRICK