Data-Centric Medical AIActive Benchmark / Project 2

License: MIT

Data-Centric AI: Evaluating MRI Preprocessing on Low-Field African Brain MRI

A Multi-Center Factorial Study of Preprocessing Impact on Model Robustness, Calibration, and Explainability

Python 3.11+PyTorch 2.0+MONAI 1.2+ANTsPySimpleITKTorchIOResNet-18DenseNet-121EfficientNet-B0

Core Research Ecosystem Vision

Evaluating whether optimizing preprocessing quality via factorial experimental design (Pipelines A–G) yields greater classification accuracy, calibration, and cross-hospital generalization than increasing model complexity on 225 low-field African brain MRI scans.

Multi-Center African Hospital Cohorts & Scanners

Private Diagnostic Centre

LifeBridge Hospital

Abuja, Nigeria

1.5T Scanner

Federal Teaching Hospital

University of Port Harcourt Teaching Hospital (UPTH)

Port Harcourt, Nigeria

0.2T - 1.5T Scanners

State Teaching Hospital

Rivers State University Teaching Hospital (RSUTH)

Port Harcourt, Nigeria

1.5T Scanner

Disease PathologySubjects (225 Total)

HydrocephalusMarked ventricular enlargement, thinning of the corpus callosum

70 subjects

DementiaGeneralized cortical atrophy, hippocampal volume loss

40 subjects

Parkinson’s DiseaseSubtle structural alterations, structural exclusion

35 subjects

EpilepsyHippocampal sclerosis, structural lesions

30 subjects

Healthy ControlsAge-matched controls, no structural abnormalities

50 subjects

Factorial Preprocessing Pipelines (Pipelines A → G)

Pipeline A (Minimal Baseline)

DICOM to NIfTI, RAS coordinate alignment, spatial resampling to 256x256x32, min-max scaling [0,1].

Pipeline B (+N4 Bias Field)

Pipeline A + N4 Bias Field Correction (ANTsPy/SimpleITK) to remove RF coil non-uniformity.

Pipeline C (+Skull Stripping)

Pipeline B + Deep learning brain extraction (HD-BET / MONAI UNet) to isolate brain parenchyma.

Pipeline D (+Intensity Z-Score)

Pipeline C + Z-score standardization (mu=0, sigma=1) & histogram matching.

Pipeline E (+NLM Denoising)

Pipeline D + Non-Local Means / Anisotropic Diffusion filtering to attenuate noise.

Pipeline F (+CLAHE Contrast)

Pipeline E + Adaptive contrast enhancement (CLAHE) for tissue-contrast definition.

Pipeline G (Complete Unified Workflow)

Full sequential pipeline: Orientation -> N4 Bias -> Brain Extract -> Denoise -> CLAHE -> Z-Score.

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/