A Novel Liver Tumor Segmentation Method Based on Multi-Branch Texture Feature Extraction

Accurate segmentation of liver tumors from Computed Tomography (CT) images remains a significant challenge due to factors such as ambiguous tumor boundaries, high inter-tumor morphological variability, and low contrast relative to surrounding hepatic parenchyma. Furthermore, prevailing deep learning segmentation networks often inadequately leverage crucial texture information inherent in CT data. To address these limitations, this paper proposes a novel automatic liver tumor segmentation method that integrates multibranch texture feature extraction. Specifically, a dedicated texture feature extraction network is strategically embedded within the skip connections of a ResUNet architecture. This design effectively captures rich, deep multi-scale texture representations characterizing liver tumors, complementing the standard spatial features learned by the main U-Net pathway. Additionally, we introduce an innovative Multi-scale Grouped Dilated Convolution (MSGDC) module within the encoder path. The MSGDC module employs parallel binary grouped convolutional layers configured with varying dilation rates, enabling efficient extraction of multiscale contextual information while expanding the network’s receptive field without significantly increasing computational complexity. This integration substantially enhances the network’s feature representation capability. Extensive experiments conducted on the public LiTS2017 liver tumor segmentation benchmark demonstrate the superior performance of our proposed method compared to several state-of-the-art approaches across key evaluation metrics. Notably, the method exhibits pronounced advantages in segmenting small tumors and regions characterized by highly blurred boundaries. These results suggest that our approach holds considerable promise as a robust technical reference for advancing automated liver tumor screening in clinical practice.