Deep learning for fully automated segmentation and volumetry of Couinaud liver segments and future liver remnants shown with CT before major hepatectomy: a validation study of a predictive model

BACKGROUND: Recent reports have shown the potential for deep learning (DL) models to automatically segment of Couinaud liver segments and future liver remnant (FLR) for liver resections. However, these studies have mainly focused on the development of the models. Existing reports lack adequate validation of these models in diverse liver conditions and thorough evaluation using clinical cases. This study thus aimed to develop and perform a spatial external validation of a DL model for the automated segmentation of Couinaud liver segments and FLR using computed tomography (CT) in various liver conditions and to apply the model prior to major hepatectomy. METHODS: This retrospective study developed a 3-dimensional (3D) U-Net model for the automated segmentation of Couinaud liver segments and FLR on contrast-enhanced portovenous phase (PVP) CT scans. Images were obtained from 170 patients from January 2018 to March 2019. First, radiologists annotated the Couinaud segmentations. Then, a 3D U-Net model was trained in Peking University First Hospital (n=170) and tested in Peking University Shenzhen Hospital (n=178) in cases with various liver conditions (n=146) and in candidates for major hepatectomy (n=32). The segmentation accuracy was evaluated using the dice similarity coefficient (DSC). Quantitative volumetry to evaluate the resectability was compared between manual and automated segmentation. RESULTS: The DSC in the test data sets 1 and 2 for segments I to VIII was 0.93±0.01, 0.94±0.01, 0.93±0.01, 0.93±0.01, 0.94±0.00, 0.95±0.00, 0.95±0.00, and 0.95±0.00, respectively. The mean automated FLR and FLR% assessments were 493.51±284.77 mL and 38.53%±19.38%, respectively. The mean manual FLR and FLR% assessments were 500.92±284.38 mL and 38.35%±19.14%, respectively, in test data sets 1 and 2. For test data set 1, when automated segmentation of the FLR% was used, 106, 23, 146, and 57 cases were categorized as candidates for a virtual major hepatectomy of types 1, 2, 3, and 4, respectively; however, when manual segmentation of the FLR% was used, 107, 23, 146, and 57 cases were categorized as candidates for a virtual major hepatectomy of types 1, 2, 3, and 4, respectively. For test data set 2, all cases were categorized as candidates for major hepatectomy when automated and manual segmentation of the FLR% was used. No significant differences in FLR assessment (P=0.50; U=185,545), FLR% assessment (P=0.82; U=188,337), or the indications for major hepatectomy were noted between automated and manual segmentation (McNemar test statistic 0.00; P>0.99). CONCLUSIONS: The DL model could be used to fully automate the segmentation of Couinaud liver segments and FLR with CT prior to major hepatectomy in an accurate and clinically practicable manner.