Attention Convolutional U-Net for Automatic Liver Tumor Segmentation

Manual radiology reporting of liver Computed Tomography (CT) volumes is time consuming process, requires expertise and is prone to human errors. Computer aided diagnosis (CAD) of liver tumor has multiple advantages including early diagnosis, treatment planning, tumor volume measurement, and monitoring clinical responses. The challenge in automatic liver segmentation in CT images is due to its heterogeneous size, shape, location of tumor, low contrast among the healthy tissues and tumorous area, and their blurred boundaries. This research focuses on automatic liver tumor segmentation using convolutional neural networks (CNN). A U-Net based attention architecture, i.e., Attention-Convolutional U-Net (ACon-U-Net) is proposed, where the attention high-resolution inception block replaces the sequence of every two convolutional layers. The model was trained, validated, and tested using the LiTS dataset. For tumor segmentation, the proposed model is evaluated in terms of dice coefficient and intersection over union. The dice score for training, validation, and testing was 97 %, 69 %, and 77 %, respectively, from abdominal CT scan images of the liver.