Class center attention network with spatial adaption for enhancing hepatic segments classification with low-visibility vascular

Abstract

Automatic classification of hepatic segments is of great use for liver surgical resection planning. However, conventional computer-aided annotation methods have difficulty annotating cases with weakly visible hepatic vascular structures. To address this issue, we proposed a class center attention convolutional neural network with spatial adaption. In our method, an improved 2.5D input module—one with spatial adaption—was used to enhance difficult to observe information on the target slices. Furthermore, a class center attention branch was introduced, which refines the unclear boundaries caused by the low-visibility hepatic vascular system by detecting the consistency of the class centers and the voxels near boundaries. Our proposed framework is trained and evaluated on 112 CT scans which were produced by the Task08_HepaticVessel in the Medical segmentation Decathlon study. Experimental results demonstrate the final algorithm for classification of hepatic segments obtained large overlaps comparable to that of manual annotations, even if the visibility of the hepatic vascular system is low.