Microaneurysm Detection for Early DR Diagnosis Using Advanced Image Processing and a Lightweight Feature Pyramid Network

Diabetic Retinopathy (DR) is a diabetes-related condition that can lead to vision loss and even permanent blindness. The growing number of diabetic patients combined with a shortage of ophthalmologists underscores the need for automated screening tools to enable early detection. Microaneurysms (MAs), the earliest signs of DR, yet identifying them in fundus images is challenging because of their tint size and subtle features. Furthermore, low contrast, noise, and lighting variations in fundus images, including glare and shadows, add complexity to the detection process. To address these challenges, we applied image enhancement techniques such as green channel extraction, gamma correction, and median filtering to improve image quality. Additionally, to boost the performance of the MA detection model, we integrated a lightweight Feature Pyramid Network (FPN) with a pre-trained ResNet34 backbone, allowing for multiscale feature capture. Our method was evaluated on the E-ophtha dataset, achieving a sensitivity of 0.601, precision of 0.697 and F1 score of 0.651. The experimental findings indicate that the proposed method surpasses current approaches.