MSCNN-AM: A Multi-Scale Convolutional Neural Network With Attention Mechanisms for Retinal Vessel Segmentation

• Main Authors: Qilong Fu, Shuqiu Li, Xin Wang
• Format: Article
• Language: English
• Published: IEEE 2020-01-01
• Series: IEEE Access
• Subjects: Retinal vessel segmentation; convolutional neural network; multi-scale information; attention mechanism
• Online Access: https://ieeexplore.ieee.org/document/9187192/

Automatic retinal vessel segmentation has drawn significant attention in early diagnosis and treatment of many diseases, such as diabetes, retinal diseases, and coronary heart disease. However, due to vessels exhibit variations in morphology and low contrast, it is still challenging to obtain accurate segmentation results. In this paper, aiming at upgrading the accuracy and sensitivity of existing vessel segmentation methods, we propose a Multi-Scale Convolutional Neural Network with Attention Mechanisms (MSCNN-AM). For extraction of blood vessels at different scales, we introduce atrous separable convolutions with varying dilation rates, which could capture global and multi-scale vessel information better. Meanwhile, in order to reduce false-positive predictions for tiny vessel pixels, we also adopt attention mechanisms so that the proposed MSCNN-AM can pay more attention to retinal vessel pixels instead of background pixels. Because the green channel shows better vessel contrast and less noise than other channels in the RGB image, our proposed MSCNN-AM is trained and tested with green channel images only, excluding extra pre-processing and post-processing steps. The proposed method is evaluated on three public datasets, including DRIVE, STARE, and CHASE_DB1. In addition, we adopt six objective metrics to verify the performance of the MSCNN-AM, including sensitivity (Se), specificity (Sp), accuracy (Acc), F1-score, an area under a receiver operating characteristic curve (AUC-ROC), and an area under precision/recall curve (AUC-PR). Experimental results indicate that our proposed method outperforms most of the existing methods with a sensitivity of 0.8342/0.8412/0.8132 and an accuracy of 0.9555/0.9658/0.9644 on DRIVE, STARE, and CHASE_DB1 separately.