Automated 3D liver segmentation from hepatobiliary phase MRI for enhanced preoperative planning.

Recent advancements in deep learning have facilitated significant progress in medical image analysis. However, there is lack of studies specifically addressing the needs of surgeons in terms of practicality and precision for surgical planning. Accurate understanding of anatomical structures, such as the liver and its intrahepatic structures, is crucial for preoperative planning from a surgeon's standpoint. This study proposes a deep learning model for automatic segmentation of liver parenchyma, vascular and biliary structures, and tumor mass in hepatobiliary phase liver MRI to improve preoperative planning and enhance patient outcomes. A total of 120 adult patients who underwent liver resection due to hepatic mass and had preoperative gadoxetic acid-enhanced MRI were included in the study. A 3D residual U-Net model was developed for automatic segmentation of liver parenchyma, tumor mass, hepatic vein (HV), portal vein (PV), and bile duct (BD). The model's performance was assessed using Dice similarity coefficient (DSC) by comparing the results with manually delineated structures. The model achieved high accuracy in segmenting liver parenchyma (DSC 0.92 ± 0.03), tumor mass (DSC 0.77 ± 0.21), hepatic vein (DSC 0.70 ± 0.05), portal vein (DSC 0.61 ± 0.03), and bile duct (DSC 0.58 ± 0.15). The study demonstrated the potential of the 3D residual U-Net model to provide a comprehensive understanding of liver anatomy and tumors for preoperative planning, potentially leading to improved surgical outcomes and increased patient safety.