Dilated Neural Networks for Improving Microaneurysm Detection in Fundus Images

Diabetic retinopathy (DR) is a microvascular complication that can lead to blindness if left untreated. Early detection of microaneurysms (MAs) is crucial for effective DR diagnosis. However, detecting MAs is challenging due to their small size, low contrast, and the presence of noise and other artifacts. This paper presents a novel deep learning approach for MA detection using a dilated residual network (DRN) with an Region of interest (ROI) layer. Unlike traditional models such as VGG16 and ResNet, our DRN model achieves a higher input resolution (28 × 28) for ROI pooling, significantly improving MA detection accuracy. Extensive experiments confirm the effectiveness of our approach, achieving superior performance compared to state-of-the-art methods on the e-ophtha-MA, DiaretDB1-MA, and ROCh datasets. Notably, our method achieves the best FROC (free-response receiver operating characteristic) score of 0.6038 on the e-ophtha-MA dataset.