Publication

Abstract : Diabetic retinopathy (DR) grading is a complex task because of the need to differentiate between subtle intraclass variations, address skewed data distributions, and identify small lesions. The solution to accurate DR grading depends on identifying specific, distinctive features that emphasize minor visual variations, such as those observed in microaneurysms, and soft exudates. However, the difficulty increases in identifying tiny and subtle abnormalities such as microaneurysms using traditional convolutional neural networks (CNNs). CNNs are spatially confined, focusing on localized regions within an image, thereby restricting their capacity to understand the global context and complex relationships between features across spatial and channel dimensions. The proposed twofold cross-feature enhancement module (2X-FEM) efficiently overcomes those limitations by enabling and facilitating cross-channel and cross-spatial interactions, boosting the network’s overall contextual understanding and feature representation. Combining spatial and channel information provides a comprehensive knowledge of images beyond the limitations of traditional CNNs. The proposed method combines cascaded dense block (CDB) and 2X-FEM to create a new architecture called CDB with a 2X-FEM (CDB-2X-FEM). This arrangement boosts the network’s capacity to extract intricate and abstract information, hence enhancing performance and the ability to recognize intricate patterns. The C2x-FNet is formed by the dense connectivity between cascaded blocks of CDB-2X-FEM, enabling efficient feature reuse. The experimental findings on three publicly accessible datasets, EyePACS, dataset for DR (DDR), and Asia Pacific Teleophthalmology Society (APTOS-2019), demonstrate exceptional performance compared with the state-of-the-art techniques.