Automatic Screening for Diabetic Retinopathy Severity

Diabetic retinopathy is a leading cause of preventable blindness, and screening at scale is limited by the number of clinicians available to read retinal images — a manual read that is also poorly reproducible between readers. Automatic severity grading offers a way to triage.

The hard version is the five-class problem — no DR, mild, moderate, severe, proliferative. Most published work at the time either collapsed it to two or three classes, or reached about 75% on five.

What made the difference

Preprocessing, more than architecture. The Kaggle dataset is a real screening archive with wildly varying image quality. Applying contrast-limited adaptive histogram equalisation to the green channel only beat applying it to all three — green carries most of the retinal vasculature contrast and the other channels mostly contribute noise. Adding a centre crop to the retina itself pushed validation accuracy from 64.65% to 67.60% in the preliminary comparison.

Progressive resizing. Train at 256 × 256, then continue at 512 × 512. This lifted EfficientNet-B4 by 3.47 percentage points of accuracy against single-size training, and B5 by 1.07 — notably, it helped the smaller model more.

Focal loss, because 25,810 of roughly 35,000 images are class 0.

Results

B5 did not beat B4 despite 1.58× the parameters, which suggests it is simply overparameterised for this dataset. At 83.89% on five-class grading, both sit well above the ~75% previously reported with larger architectures.

Carried out at the Biomedical Research Group, Bandung Institute of Technology, and published at IEEE TENCON 2020.