An Enhanced U-Net Model with Local and Global Context for Liver Segmentation from CT Images

Liver segmentation is vital for diagnosing liver diseases, planning surgeries, and monitoring treatment. Deep learning techniques like Convolutional Neural Networks (CNNs) and U-Net have shown satisfactory results in liver segmentation tasks. However, challenges such as irregular liver shapes, small liver size, and fuzzy liver boundaries significantly impact the performance of CNNs and U-Net. To address these challenges, leveraging contextual information has been proven beneficial. In this paper, the proposed method is an end-to-end automatic liver segmentation framework that incorporates two innovative modules: the Adaptive Spatial-Channel Convolution (ASCC) module and the Enhanced Fusion Attention (EFA) module. The ASCC module extracts local contextual information while reducing spatial and channel redundancy to enhance feature representation. The EFA module effectively combines local contextual information from the Convolutional Neural Network with global contextual information from the Transformer, resulting in a more comprehensive contextual understanding. We thoroughly evaluated our method on the LiTS dataset and achieved an impressive Dice similarity coefficient (DSC) of 0.9568 for liver segmentation. The experimental results unequivocally demonstrate the highly satisfactory segmentation performance of our proposed method.