A Lesion-Aware Detection Framework for Red Lesions in Retinal Fundus Images

Diabetic retinopathy (DR) is a serious eye condition caused by damage to the retina due to prolonged diabetes. Among the first signs of DR are microaneurysms and hemorrhages, collectively called red lesions. Detecting these lesions is crucial in preventing vision loss. This paper presents a refined feature attention (RFA) module that refines features from the feature pyramid network (FPN) layers before they are passed to the detection head. Dataset-specific adaptive anchor ratios are derived using a quantile-based multi-stage clustering strategy to better align the anchors with lesion geometry. A lesion-aware sampling strategy based on the morphological characteristics of red lesions is proposed to balance the batch during training for datasets with a wide range of spatial distributions and structural complexities. The network's performance was evaluated across three standard retinal image datasets: Messidor-A, SaNMoD, and E-Ophtha MA. On the Messidor-A and SaNMoD datasets, the proposed method achieved FROC scores of 0.4650 and 0.4429, respectively. Furthermore, evaluation on the E-Ophtha MA dataset yielded an FROC score of 0.549, demonstrating the robustness and generalization capability of the proposed model. The proposed network achieved a precision of 73.6%, a recall of 69.0%, and an F1-score of 71.2% on the Messidor-A dataset, with false positives per image (FPI) of 2.802 and an average inference time of 45.24 ms per image. The model requires 128 GFLOPs and has 36.35 million parameters. Experiments show that combining the attention module with adaptive anchoring and lesion-aware sampling improves the model's ability to detect red lesions.