Dual attention U-net for liver tumor segmentation in CT images

Abstract

Segmenting liver tumors in CT scans plays a vital role in medical analysis planning. The clinicians require a detailed 3D understanding of the tumor’s location and liver anatomy, to decide about the proper surgical resection approach. Manual segmentation requires a lot of efforts and time also it depends on the expertise of clinicians. An automatic U-net based method for liver tumors delineation in CT images is proposed. It relies on employing attention-based processes to enhance the performance of U-net. Hard attention and soft attention are used to orient the U-net in learning the intended features from the target CT scans. Soft attention mechanisms, spatial and channel attentions, are employed to help in extracting the long-range relationships and allow the network to successfully distinguish tumors from the surrounding parenchyma. The paper addressed the use of region based active contour technique as postprocessing step to improve the predicted segmented tumors. The proposed approach is validated using a challenging big LiTS datasets. The achieved Dice score for the segmenting of liver tumors is 0.81. The suggested method was successful in discriminating liver tumors from surrounding tissue in heterogeneous CT scans, demonstrating its generalizability and reliability to be used for automatic analysis of the liver tumors in daily clinical practice.