African Brain MRI AI Research Ecosystem

A Multi-Project Research Program for Robust, Explainable & Clinically Validated AI in African Neuroimaging

PythonPyTorchMONAISimpleITKDockerafri_brain_coreBIDSBayesian Models

Core Research Ecosystem Vision

Documents the architecture, dependency matrix, and 7-phase roadmap of a comprehensive research ecosystem designed to bridge the gap between idealized research MRI and real-world low-resource African clinical environments.

7-Phase Ecosystem Project Roadmap

Phase 1: Foundation Infrastructure

Project 1: African Clinical Brain MRI Benchmark

Establish the foundational benchmark ecosystem for evaluating AI models on heterogeneous African clinical MRI data.

Contributions: Multi-centre Nigerian dataset, standardized preprocessing, patient-level splits, ML & deep learning baselines, calibration & robustness testing.

Project 2: Data-Centric AI for Low-Field MRI

Investigate how data quality, intensity normalization, augmentation, and preprocessing influence AI performance in low-field MRI.

Contributions: Optimized preprocessing recommendations, data quality framework, and low-field MRI AI best practices.

Phase 2: Biological & Statistical Understanding

Project 3: Bayesian Hierarchical Modeling of Brain Morphometry

Develop uncertainty-aware statistical models for structural brain differences across neurological disorders in Nigerian populations.

Contributions: Population-specific brain biomarkers, uncertainty-aware inference, site variance estimation, and disease-specific morphometric patterns.

Phase 3: AI Reliability Evaluation

Project 4: Robust Cross-Hospital Generalization

Evaluate whether AI models trained in one clinical environment safely generalize across different scanners and hospital protocols.

Contributions: Leave-one-hospital-out validation, domain shift analysis, and deployment readiness framework.

Project 5: Quantifying MRI Image Quality Effects on AI Performance

Measure how progressive SNR, blur, sharpness, and motion artifacts affect AI classification reliability boundaries.

Contributions: AI failure thresholds, quality-performance relationships, and operational safety limits.

Project 6: Explainability and Clinical Plausibility

Audit whether AI models learn true anatomical pathology or misleading scanner-specific shortcuts using Grad-CAM & saliency stability.

Contributions: Grad-CAM localization, anatomical consistency, saliency stability, and clinical plausibility evaluation.

Phase 4: Clinical Translation

Project 7: DeepAccess-MRI

Develop AI-assisted MRI protocol optimization through clinically validated missing sequence reconstruction to reduce scan time.

Contributions: Sequence reconstruction, transfer learning, uncertainty estimation, pathology preservation testing, and clinical validation.

Shared Core Architecture (afri_brain_core)

├──preprocessing/ — N4 bias correction, resampling, reorientation

├──datasets/ — Subject-level PyTorch Dataset loaders for T1, T2, FLAIR

├──evaluation/ — Calibration (ECE), RRI, AURC, and DeLong statistical tests

├──quality_control/ — MRIQC signal-to-noise (SNR), CNR, EFC, and FWHM profiling

├──explainability/ — Grad-CAM localization & attribution stability

├──morphometry/ — Automated segmentation & subcortical nuclei volume extraction

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