Attention-Based Deep Fusion Network for Retinal Lesion Segmentation in Fundus Image

Abstract

Segmentation of subtle lesions in fundus images has become a vital part of diagnosing ocular diseases such as Diabetic Retinopathy (DR). Diabetic eye disease is characterized by the scattered lesions in the retina. Detection of these lesions at the early stage is important as its progression leads to vision loss if proper treatment is not taken. The main objective of the work is to assist ophthalmologist in the effective diagnosis of eye disease providing timely treatment. This paper focuses on developing a deep learning-based Fusion Network (Fu-Net) with an attention mechanism for lesion segmentation in color fundus images. The network was developed based on the baseline U-Net model with trivial modification in the encoder and decoder part of the model. A multi-feature fusion block (MFuse) is integrated with the encoder of the network to extract the lesion features and a channel attention module is integrated with the decoder part to fuse the feature information effectively. Besides, a modified weighted focal loss function is introduced to mitigate the problem of class imbalance in the fundus image. The computational results obtained signifies the superior performance of the proposed method in the lesion segmentation task.