Abstract
BACKGROUND: To develop and validate a deep learning-based future liver remnant (FLR) volumetry system (DL-FLRVS) by secondary utilization of existing preoperative CT images, and to compare the differences between the currently used method and DL-FLRVS in volumetry of FLR and the application in candidate categorization before major hepatectomy. METHODS: DL-FLRVS, which consists of five 3D U-Net models for the automated segmentation of liver anatomy on contrast-enhanced CT, was developed (n = 307, 170, 170, 170, and 492, respectively) and validated (n = 178) in external validation cohorts. The FLR and FLR% of patients were measured using DL-FLRVS and the hepatic VCAR (i.e., a semi-automated segmentation program on a dedicated workstation) for different types of major hepatectomy. Manual measurements were used as a reference. The differences in FLR assessment and candidate categorization between the two methods were compared using Spearman analysis and McNemar’s test, respectively. RESULTS: The mean FLR and FLR% values were (493.51 ± 284.77) cm(3) and (38.53 ± 19.38) %, respectively, when DL-FLRVS was used, (489.23 ± 286.29) cm(3) and (37.77 ± 19.12) %, when the hepatic VCAR was used. No significant differences in the categorization of candidates for major hepatectomy were found between the DL-FLRVS and human doctors (P > 0.99 and P > 0.99, respectively) or between the hepatic VCAR and human doctors (P > 0.99 and P = 0.07, respectively). CONCLUSION: DL-FLRVS represents a potential alternative to the currently used method in volumetry of FLR and FLR-based candidate categorization in major hepatectomy in clinical practice. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12880-025-02106-0.