yashvmaurya/glaucoma-pipeline

Glaucoma Two-Stage AI Pipeline

Artifacts for a two-stage glaucoma decision pipeline.

• Stage 1 — Detect: EfficientNet-B0 fine-tuned on SMDG-19 fundus images (stage1_detect/stage1_best.pt) with Grad-CAM explainability. External test on REFUGE1.
• Stage 2 — Progression (novel core): GRU residual forecaster of the next visual-field test from history, with calibrated MC-dropout uncertainty (stage2_progress/stage2_gru.pt).

Stage 2 verified results (UWHVF, patient-level test set)

• Forecasting: GRU 2.74 dB MAE < last-value 2.78 < pointwise-LR 3.28.
• Calibrated uncertainty: ECE 0.048, 90% interval coverage 91.7%.
• Severity staging: accuracy 0.815, macro-F1 0.743. Fast/slow (MD-slope): AUROC 0.855.

Trained models are released for research/reproducibility. Datasets are public (SMDG-19, REFUGE, UWHVF) and are not redistributed here. Not a medical device.

See the project repo for code, config, and the full reliability report.