1. Input
APTOS retinal fundus images with diabetic retinopathy labels from grade 0 to grade 4.
Pipeline
End-to-end workflow
2. Preprocessing
Image resizing, normalization, augmentation, and stratified train-validation-test partitioning.
3. Modeling
Transfer learning with EfficientNetB0, ResNet152V2, or NASNetLarge plus a Random Forest baseline.
4. Evaluation
Accuracy, confusion matrix, classification report, training logs, saved models, and JSON summaries.
Comparison
Architecture comparison surface
The cards below compare the role and tradeoffs of each model family used in the project.
Results
What the project produces
This presentation layer does not invent experimental numbers. It shows the exact evaluation outputs the codebase is built to generate on a final run.
Implementation
What was refined for the final version
Path cleanup
The project was decoupled from Kaggle-only file paths and adapted to a clean local repository layout.
Unified entrypoint
The codebase now exposes one final CLI entrypoint with backbone selection, smoke testing, and artifact export.
Presentation layer
README, architecture diagram, and this GitHub Pages UI were added so the project reads like a product, not a notebook dump.
Future Prospect