Enhancing Liver Segmentation: A Deep Learning Approach with EAS Feature Extraction and Multi-Scale Fusion

Abstract

Deep learning technology have been broadly used in segmentation tasks of liver. To address the limitation of suboptimal segmentation for small targets, an end-to-end EAS(ECA-Attention and Separable convolution) U-Net is proposed based on deep learning. The basic module employs depthwise separable convolutional modules instead of convolutional modules to reduce the parameters count and enhance the extraction of deep-layer information. The pyramid module based on Efficient Channel Attention (ECA) is utilized to obtain different receptive fields. And that model can overcome the limitation of the U-Net model with a single receptive field and improve the segmentation capability for targets of different sizes. A deep supervision module with multi-scale output fusion is designed to extract detailed information about liver with high quality. The proposed method is tested on the Liver Tumor Segmentation (LiTS) dataset for liver segmentation, achieving a Dice Similarity Coefficient (DSC) of 92.20% for liver segmentation. Compared to existing models, the proposed method demonstrates higher accuracy in liver segmentation.