Publication

Abstract : Diabetic Maculopathy (DM), a serious complication of diabetes, can lead to permanent vision loss if not detected and treated early. Accurate identification of diabetic retinopathy (DR) related features, such as microaneurysms (MA), exudates (EX), and hemorrhages (HE), is critical for timely intervention. Traditional methods often miss these subtle features, delaying early diagnosis. This paper propose a novel self-calibrated convolutional neural network designed for fine-grained classification of key retinopathy features viz., DiabetNet. By replacing standard convolutional layers with self-calibrated blocks, DiabetNet enhances feature extraction and inter-layer information flow. Additionally, a progressive transformer structure is integrated to capture global dependencies, further improving the detection process. DiabetNet's performance is benchmarked against CNN architectures like ResNet-18, ResNet-50, and EfficientNet. The results demonstrate significant improvements in sensitivity, specificity, and accuracy for detecting MA (70.7%), EX (76.76%), and HE (77.64%). Notably, DiabetNet achieves these results using only color fundus photographs (CFP), eliminating the need for fluorescein angiography fundus images (FAF), and thus reducing the cost and complexity of diagnosis. Thus, DiabetNet offers a highly accurate, cost-effective solution for the early detection of DM, aiding in the prevention of vision loss in Diabetic Patients.