Diabetic Retinopathy Screening
Explainable Grading for Rural Screening Programmes
Captured from the live deployment, not a mockup
01 · The problem
Grading a 0 as a 4 is not the same mistake as grading a 0 as a 1, so trained ICDR grades 0 to 4 with rank-consistent ordinal heads instead of plain cross-entropy, on patient-level splits where both eyes never straddle train and test.
02 · How it works
- 01
Gated every image through a 3-class gradability model fused with focus, illumination and field-of-view features, returning why an image was rejected so the operator can recapture rather than guess.
03 · What it cost, and what it returned
Detected sub-pixel microaneurysms with morphological top-hat and matched filtering before CNN classification, and measured Grad-CAM localisation against annotated lesions rather than asserting the heatmaps meant anything.