How rajkumarprt/multi-class-diabetic-retinopathy-detection-using-xai-driven-cnn-transformer-fusion-framework works
This repository trains retinal-image classifiers and serves diabetic-retinopathy predictions with visual explanations and model-comparison reporting. Its main workflow is OCT2017 CNN pretraining, APTOS training and QWK-based model selection, then web inference with Grad-CAM/attention explanations. The README documents the application’s comparison, results, and reference-gallery capabilities; the sampled source does not show every route or UI data path, so those links are documented rather than inferred from folder names.
Main entry points: train.py Training pipeline · app.py Flask inference app · dataset.py Dataset loaders · explain.py Prediction explanations
Components 20
Data preparation
- Fundus datasets
- Preprocessing
data/preprocessing.py - Dataset loaders
data/dataset.py
Training and evaluation
- Training pipeline
train.py - Runtime configuration
config.py - Train and evaluate
utils/engine.py - Evaluation metrics
utils/metrics.py - CNN classifier
models/cnn_model.py - ViT classifier
models/vit_model.py - Fusion classifier
models/fusion_model.py - Selected weights
checkpoints - Training reports
reports/generate_report.py - Benchmark report files
outputs/reports
Inference and explainability
- Flask inference app
webapp/app.py - Prediction explanations
explainability/explain.py - Grad-CAM
explainability/gradcam.py - Attention rollout
explainability/attention_rollout.py
Web reports
- Web report exports
webapp/report_generator.py
Other components
- Image submitter
- Web application user
Connections 23
- Fundus datasets → Dataset loaders: loads images
data/dataset.py - Dataset loaders → Preprocessing: preprocesses images
data/dataset.py - Training pipeline → Dataset loaders: builds loaders
train.py - Training pipeline → Runtime configuration: uses settings
train.py - Training pipeline → CNN classifier: trains CNN
train.py - Training pipeline → ViT classifier: trains ViT
train.py - Training pipeline → Fusion classifier: trains fusion
train.py - Training pipeline → Train and evaluate: runs training
train.py - Train and evaluate → Evaluation metrics: computes metrics
utils/engine.py - Training pipeline → Selected weights: selects best model
README.md - Training pipeline → Training reports: writes results
train.py - Training reports → Benchmark report files: writes reports
reports/generate_report.py - Training pipeline → Dataset loaders: pretrains on OCT2017
train.py - Image submitter → Flask inference app: uploads image
README.md - Flask inference app → Prediction explanations: requests explanation
webapp/app.py - Prediction explanations → Grad-CAM: generates heatmap
explainability/explain.py - Prediction explanations → Attention rollout: computes attention
explainability/explain.py - Flask inference app → CNN classifier: compares models
README.md - Flask inference app → ViT classifier: compares models
README.md - Flask inference app → Fusion classifier: compares models
README.md - Web application user → Flask inference app: views results
README.md - Flask inference app → Web report exports: serves report exports
webapp/app.py - Web report exports → Benchmark report files: reads metrics
webapp/report_generator.py