DilsaraAbeywarana/glaucoma-severity-system

Glaucoma Severity System

Three-stage system for retinal fundus photographs: input validation, severity grading, and quantitative cup-to-disc measurement.

Part of the NeuroGSD project (AI-based early detection of common eye diseases).

Scope

This module grades severity in confirmed glaucoma cases. It does not screen for the presence of glaucoma. A healthy fundus photograph will still receive a severity grade. Screening is not implemented.

Usage

from inference import GlaucomaSystem
import cv2

sys = GlaucomaSystem(".")
result = sys.predict(cv2.imread("fundus.jpg"))

Accepted input returns grade, confidence, class probs, vcdr and cup_disc_area_ratio. Rejected input returns accepted: False with the reasons and produces no grade.

Components

1. Input validation — two independent checks on the raw image, before any preprocessing. Order matters: the disc-cropping stage transforms an arbitrary photograph into a texture patch, destroying the evidence needed to reject it. The first check uses black-surround fraction and red-channel dominance; the second measures Mahalanobis distance from the training distribution in EfficientNetB4 feature space. Photographs of a car and of a flower passed all statistical checks and were caught only by the distance measure, so both are needed.

2. Severity classifier — EfficientNetB4 fine-tuned on optic disc crops, three-way softmax over mild / moderate / severe. Preprocessing: black-border trim, disc localisation by brightest-point detection with a 12% border exclusion, crop at 40% of the shorter dimension, CLAHE (clip 2.0, 8x8, LAB L-channel), resize to 384.

3. Cup-to-disc measurement — U-Net trained on ORIGA, REFUGE and G1020 (2,237 image-mask pairs), returning vertical CDR and cup-to-disc area ratio.

Results

Severity classification, held-out test set (n=67), evaluated once:

Metric	Value
Quadratic weighted kappa	0.604
Accuracy	0.672
Macro-F1	0.660
Within-one-grade agreement	0.940

Cross-validated development estimate was 0.607, closely matching the held-out result and indicating a leakage-free evaluation.

Segmentation, held-out REFUGE test partition (a source not seen during training): rim Dice 0.883, cup Dice 0.862.

Input validation: 93% of fundus images from an unseen camera accepted; all non-fundus images tested rejected.

Limitations

• Trained on 322 images from a single clinic; generalisation untested
• Labels student-graded under expert guidance, not independently verified
• No healthy class; cannot screen for the presence of glaucoma
• Segmentation quality on clinic images is lower than on the public validation set, since their disc crops are looser than the training crops. Measured CDR correlates with expert grade (Spearman 0.219 vertical, 0.309 area) but this survives on clean masks only at 0.194 and 0.282
• Research use only. Not a medical device.