An Explainable Framework Using Fundus Image Preprocessing and Grad-CAM++

Warning: This code is for research and educational purposes only. Any clinical deployment requires IRB approval and prospective field validation.

Fundus image classifiers must achieve accuracy and efficiency while remaining clinically interpretable. This repository provides a reproducible pipeline on APTOS 2019 that combines (i) robust preprocessing (circular masking, CLAHE, green-channel emphasis), (ii) lightweight EfficientNet models, and (iii) Grad-CAM++ visual explanations.

• Does strengthened preprocessing improve performance and stability on DR screening?
• Can lightweight models (EfficientNet-B0/B3) reach near-clinical screening performance?
• Do Grad-CAM++ visualizations align with pathologic regions used in clinical reading?

• Validation: AUROC 0.9800, AUPRC 0.9689
• Test: AUROC 0.9855, AUPRC 0.9780
• Threshold at specificity ≈ 0.95: thr = 0.467 → Spec 0.940, Sens 0.933
• Grad-CAM++ highlights optic disc, vessel branching, and hemorrhage/exudate regions consistent with clinical findings.

• APTOS 2019 Blindness Detection (3,662 images). https://www.kaggle.com/competitions/aptos2019-blindness-detection
• Reformulated as binary: non-referable (0–1) vs referable DR (≥2).

1. Circular masking (retina extraction)
2. CLAHE contrast enhancement
3. Green-channel emphasis
4. Resize to 448×448

• Backbones: EfficientNet-B0/B3 (ImageNet pretrained)
• Optimizer: AdamW (lr=3e-4, weight_decay=1e-4)
• Scheduler: Cosine LR

• AUROC, AUPRC
• Sensitivity at fixed specificity ≈ 0.95 (validation-optimized threshold)

• Grad-CAM++ on the last conv block; overlay boards for validation cases.

Figure 1. Validation: Original fundus (left of each pair) and Grad-CAM++ overlay (right).

The model consistently attends to clinically relevant structures—optic disc, vessel bifurcations, and lesion regions (hemorrhages/exudates)—supporting interpretability of predictions.

dr-xai/
├── configs/                 # experiment configs
├── src/
│   ├── dataset.py           # loader / splits / labels
│   ├── preprocess.py        # circular mask, CLAHE, green channel
│   ├── model.py             # EfficientNet wrappers
│   ├── train.py             # training loop (AMP optional)
│   ├── evaluate.py          # ROC/PR, threshold tuning, metrics
│   └── cam.py               # Grad-CAM++ generation and boards
├── scripts/                 # CLI utilities (boards, exports, end-to-end)
├── notebooks/               # Kaggle-friendly reproduction
├── artifacts/
│   ├── figs/                # test_roc.png, test_pr.png, metrics_summary.csv
│   └── cams_resume/         # val_*.png (original vs CAM overlays)
└── README.md

• Python 3.10+
• PyTorch ≥ 2.1, timm, albumentations, opencv-python, scikit-learn, matplotlib, pandas
• Grad-CAM++: pytorch-grad-cam

pip install torch torchvision torchaudio
pip install timm albumentations opencv-python scikit-learn matplotlib pandas tqdm pytorch-grad-cam

python -m src.train \
  --config configs/aptos.yaml \
  --model effnet_b0 \
  --img-size 448 \
  --epochs 3 \
  --batch-size 32

python -m src.evaluate \
  --ckpt runs/effb0_best.pt \
  --img-size 448 \
  --fix-spec 0.95 \
  --out-dir artifacts/figs

python -m src.cam \
  --ckpt runs/effb0_best.pt \
  --split val \
  --n 12 \
  --out-dir artifacts/cams_resume

• artifacts/figs/test_roc.png
• artifacts/figs/test_pr.png
• artifacts/cams_resume/val_*.png (original vs CAM overlays)

[Figure 1] Validation: Original Fundus vs CAM Overlay — Grad-CAM++ shows attention over optic disc, vessel bifurcations, and lesion areas, supporting clinical interpretability.

• External validation on EyePACS and Messidor-2 is required to assess generalization.
• Extension to multi-class DR severity.
• Latency & memory profiling for mobile/edge deployment.

• Gulshan V, et al. JAMA, 2016. doi:10.1001/jama.2016.17216
• Chattopadhay A, et al. WACV, 2018. https://doi.org/10.48550/arXiv.1710.11063
• Tan M, Le Q. EfficientNet, ICML 2019
• APTOS 2019 Blindness Detection Dataset (Kaggle)

Diabetic Retinopathy, Fundus Imaging, Deep Learning, Explainable AI, Preprocessing, Grad-CAM++

• Code: MIT (update as appropriate).
• Dataset: follow APTOS 2019 license and terms; data is not bundled.
• This project is a decision-support research tool, not an autonomous diagnostic system. Deployment requires independent validation, human oversight, conservative thresholds, ongoing monitoring, and rollback procedures.