(M)SLAe-Net: Multi-Scale Multi-Level Attention embedded Network for Retinal Vessel Segmentation

TL;DR

This paper introduces (M)SLAe-Net, a multi-scale, multi-level attention CNN for retinal vessel segmentation that effectively captures local and global features, improving accuracy and robustness across multiple datasets.

Contribution

The paper presents a novel multi-scale, multi-level attention network with a dynamic dilated pyramid pooling module and a specialized attention mechanism, reducing the need for multi-stage processing.

Findings

Outperforms existing methods on DRIVE, STARE, HRF, and CHASE-DB1 datasets.

Effectively segments vessels with varying shapes and sizes.

Demonstrates improved cross-dataset generalization.

Abstract

Segmentation plays a crucial role in diagnosis. Studying the retinal vasculatures from fundus images help identify early signs of many crucial illnesses such as diabetic retinopathy. Due to the varying shape, size, and patterns of retinal vessels, along with artefacts and noises in fundus images, no one-stage method can accurately segment retinal vessels. In this work, we propose a multi-scale, multi-level attention embedded CNN architecture ((M)SLAe-Net) to address the issue of multi-stage processing for robust and precise segmentation of retinal vessels. We do this by extracting features at multiple scales and multiple levels of the network, enabling our model to holistically extracts the local and global features. Multi-scale features are extracted using our novel dynamic dilated pyramid pooling (D-DPP) module. We also aggregate the features from all the network levels. These effectively resolved the issues of varying shapes and artefacts and hence the need for multiple stages. To assist in better pixel-level classification, we use the Squeeze and Attention(SA) module, a smartly adapted version of the Squeeze and Excitation(SE) module for segmentation tasks in our network to facilitate pixel-group attention. Our unique network design and novel D-DPP module with efficient task-specific loss function for thin vessels enabled our model for better cross data performance. Exhaustive experimental results on DRIVE, STARE, HRF, and CHASE-DB1 show the superiority of our method.