Liver Tumor Segmentation using Hybrid Residual Network and Conditional Random Fields

Liver cancer is a leading cause to cancer-related death. The accurate delineation of the liver and its abnormalities is of utmost significance in the clinical interpretation and therapeutic strategizing of hepatic diseases. The primary objective of this study is to employ a hybrid residual network model, based on the U-Net architecture, for the automated segmentation of liver and liver cancers from abdominal Computed Tomography (CT) images. The U-shaped design and skip connections of the U-Net architecture make it particularly effective for CT image segmentation by combining high-level contextual information from the contracting path with fine-grained spatial features from the skip connections. Residual Networks (ResNets) utilize residual blocks to facilitate the construction of deep neural networks, therefore addressing the challenges posed by the vanishing gradient problem. The proposed methodology yielded dice global scores of 0.94 and 0.73 for liver segmentation and liver tumor segmentation, respectively on LiTS dataset. The findings demonstrate the exceptional performance and efficacy of the approach. The methodology also incorporates histogram equalization as a preprocessing strategy and utilizes conditional random fields as a postprocessing technique in order to improve the accuracy of segmentation.