Diabetic retinopathy across datasets

Architecture

• A cross-dataset protocol: training exclusively on EyePACS and evaluating on Messidor-2, with no adaptation. No Messidor-2 image takes part in training or in checkpoint selection.
• Labels binarized by the same rule on both datasets: grade 0 of the severity scale becomes the negative class, grades 1 to 4 the positive class, which matches the refer or do-not-refer decision.
• A within-dataset split grouped by patient: 17,563 EyePACS patients, two eyes each, split into 14,050 for training and 3,513 for validation, with zero patient overlap and 26.5% positives on both sides.
• Three preprocessing strategies compared: plain resizing as the baseline, local contrast enhancement with CLAHE on the L channel of LAB space, and green channel extraction replicated across the three channels.
• Two architectures of different capacity: a SimpleCNN with 1,146,561 parameters trained from scratch, with a 36 by 36 pixel receptive field, and an EfficientNet-B0 with 4,008,829 parameters initialized from ImageNet, fine-tuned end to end.
• An identical training recipe across variants: BCEWithLogitsLoss with pos_weight 2.77, AdamW, CosineAnnealingLR, an effective batch of 32 through gradient accumulation, mixed precision, and augmentation by flips, rotation and brightness jitter.
• Metrics with 95% confidence intervals from a 2,000-replicate bootstrap: AUC-ROC, sensitivity, specificity, precision, F1 and the confusion matrix.
• Two operating points reported: the default 0.5 threshold and a threshold calibrated on validation for a minimum sensitivity of 0.90, then transferred to the external dataset with no readjustment.
• Grad-CAM implemented with hooks on the last convolutional layer, shared between the monograph figure and a Gradio web demo that loads the six checkpoints.

Implementation details

• The filter runs at native resolution and only then does the image go down to 512 by 512, because CLAHE depends on neighbourhood statistics and a microaneurysm covers few pixels in the original.
• Confidence intervals come from a bootstrap that resamples the 874 exams rather than loose images, because 86.3% of eye pairs carry concordant labels.
• The best checkpoint is picked by EyePACS validation AUC, never by the external dataset, and AUC is computed by a Mann-Whitney U statistic implemented in the project itself.