MSGNet: Modified MobileNet-ShuffleNet-GhostNet Network for Lightweight Retinal Vessel Segmentation

When it comes to the early detection of eye-related ailments like diabetic retinopathy and hypertensive retinopathy, the accurate segmentation of retinal blood vessels is of the utmost importance. However, most of the time, the segmentation model requires a computationally expensive machine, which is burdensome when it has to be employed in a remote area with a limited supply of power sources and internet access. Hence, it is necessary to design a lightweight model with performance commensurate with the state-of-the-art (SOTA) models. In this work, we present MSGNet (Modified MobileNet-ShuffleNet-GhostNet Network), a lightweight network for retinal vessel segmentation. MSGNet is a lightweight U-shaped retinal vessel segmentation model that leverages several components from lightweight architectures, i.e., depthwise separable convolution, channel shuffle operation, and the Ghost module. Then, batch normalization inside each of the components is replaced with instance normalization, which encourages cross-dataset generalization capability. Computationally speaking, MSGNet runs fast on the CPU. MSGNet is trained with the Digital Retinal Images for Vascular Extraction (DRIVE) dataset and validated using the Child Heart and Health Study (CHASE DB1) dataset and the Structured Analysis of the Retina (STARE) dataset. MSGNet shows comparable performance against the SOTA models and has the ability to generalize across different retinal vessel segmentation datasets.