EFENet: Edge and feature enhancement network for stroke lesion segmentation

Abstract

Automatic lesion segmentation holds significant clinical value for the diagnosis and rehabilitation of brain stroke. However, it faces challenges such as blurred lesion edges and variations of lesion morphology and size. To tackle these problems, we propose a model for stroke lesion segmentation based on edge and feature enhancement named EFENet. First, we propose an edge-aware decoder (EAD), which first predicts the overall lesion region and then extracts the predicted edges using a morphology-based method. An edge feature enhancement component is incorporated in the EAD to strengthen the lesion edge features in the image, alleviating the impact of edge blurring on segmentation performance. Second, local-enhanced Swin Transformer (LE-Swin) blocks are introduced in the encoders. A convolution-based local feature extraction branch is added to the window-based multi-head self-attention (W-MSA), enhancing the model’s ability to capture both global and local features. Finally, a Channel Attention Fusion module (CAF) is employed at skip connections to fuse the encoder’s global features and the decoder’s edge-enhanced features using channel attention, reducing the feature gap. Extensive experiments are conducted on two public datasets, ATLAS and ISLES2022. EFENet achieves Dice coefficients of 0.5906 and 0.7598, respectively.