LCAM-Net: Local Context Attention Network for Diabetic Retinopathy Severity Classification

Diabetic Retinopathy (DR) is a chronic condition caused by microvascular complications of diabetes mellitus. DR is the leading cause of blindness in working-age adults globally. Hence, early diagnosis and precise treatment are paramount to prevent visual impairment. However, the task’s difficulty and the high demand for ophthalmological services often lead to delays and inaccuracies in DR severity assessment. Deep Learning (DL) techniques have emerged as a promising solution to address this challenge and provide timely and unbiased DR analysis. This paper proposes a novel local-context attention mechanism (LCAM) to enhance the identification of salient elements such as lesions in DR severity classification. LCAM leverages spatial context information to emphasize abnormal regions in intermediate feature maps of a convolutional neural network by learning context weights to integrate multi-level features effectively. Experimental results on the benchmark dataset IDRiD show that incorporating the proposed module over five backbone classification architectures enhances performance, achieving state-of-the-art results: 64.08% for Quadratic Weighted Kappa score and 69.13% for Accuracy. Additionally, the LCAM-based architectures show explainability qualities to highlight lesion-related areas in the fundus image, as evidenced by using Grad-CAM.