Transforming diabetic retinopathy detection with deep learning and handcrafted feature fusion

Abstract

The task of early and accurate screening of diabetic retinopathy (DR) from fundus images is affected by the presence of class imbalance, small grade differences, and the lack of interpretability of deep learning models. This paper proposes HandEffiTrans-DR, a hybrid model that represents lesion statistics as semantic tokens and combines them with EfficientNetB0 deep features using a multi-head transformer encoder. A stepwise grading approach for five-class, three-class, and binary classification allows for progressive improvement of decision boundaries while keeping the grading relevant. Hand-engineered haemorrhage features, tested for quantitative trend and lesion map validation, help focus attention on pathology-defined areas during fusion. Experiments on the Kaggle EyePACS dataset report accuracies of 0.81, 0.88, and 0.98 for the three stages, with an area under the curve (AUC) of 1.00 for binary screening. Baseline comparisons clearly indicate the effectiveness of the proposed hybrid model over hand engineered and Convolutional neural network (CNN)-based models. Cross dataset testing on Asia Pacific Tele-Ophthalmology Society (APTOS) and an independent EyePACS dataset sustains accuracy above 0.95, ensuring the effectiveness of the proposed model.