Automatic CT liver Couinaud segmentation based on key bifurcation detection with attentive residual hourglass-based cascaded network. (May 2022)

• Record Type: Journal Article
• Authors: Wang, Manyang; Jin, Renchao; Lu, Jiayi; Song, Enmin; Ma, Guangzhi
• Is Part Of: Computers in biology and medicine. Volume 144(2022)
• Publication Date: 2022-05
• Subjects: Key point detection -- Couinaud segmentation -- Attentive residual mechanism -- Hourglass network; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2022.105363
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: This paper presents an automatic Couinaud segmentation method based on deep learning of key point detection. Assuming that the liver mask has been extracted, the proposed method can automatically divide the liver into eight anatomical segments according to Couinaud's definition. Firstly, an attentive residual hourglass-based cascaded network (ARH-CNet) is proposed to identify six key bifurcation points of the hepatic vascular system. Subsequently, the detected points are used to derive the planes that divide the liver into different functional units, and the caudate lobe is segmented slice-by-slice based on the circles defined by the detected points. We comprehensively evaluate our method on a public dataset from MICCAI 2018. Experiments firstly demonstrate the effectiveness of our landmark detection network ARH-CNet, which is superior to that of two baseline methods, also robust to noisy data. The average error distance of all predicted key points is 4.68 ± 3.17 mm, and the average accuracy of all points is 90% with the detection error distance of 7 mm. We also verify that summation of the corresponding heat-maps can improve the accuracy of point localization. Furthermore, the overlap-based accuracy and the Dice score of our landmark-derived Couinaud segmentation are respectively 91% and 84%, which are better than the performance of the direct segmentation approach and the traditional plane-based method, thus our method can be regarded as a good alternative for automatic Couinaud segmentation. Highlights: This work presents an automatic CT Couinaud segmentation method based on deep learning of keypoint detection, avoiding the complex operations on the vascular system and saving a lot of time for entire classification process. The designed network for automatic landmark localization contains multiple HG stages with attention mechanisms and intermediate supervision. This novel structure helps to pay special attention to the features near key points. Only six clearly defined bifurcations of intrahepatic vascular system need to be manually labelled for each sample when preparing training dataset, rather than voxel-by-voxel labeling as in direct semantic segmentation methods. A variant of focal loss is designed to further solve the problem of imbalance between positive and negative samples, where the neighbors of keypoints are also regarded as keypoints.