Automatic Liver Vessels Segmentation Using ResDense UNet And An Appropriate Preprocessing Pipeline

The 3D reconstruction of hepatic vessels has raised relevant clinical interest in various surgical procedures, including preoperative planning. This paper explores a novel deep-learning architecture for automatically reconstructing digital twins of hepatic vessels by comparing it with traditional 3D-UNet and 3D-VNet methods. Our approach automatically recognizes vessels from the portal vein (portal vessels) and the cava vein (hepatic vessels), employing innovative preprocessing techniques, introducing a fitted Hounsfield unit (HU) range, and strategically avoiding resampling the entire dataset to a common spacing. These innovations resulted in a more continuous representation of vascular structures and reduced outliers voxels. The 3D architecture used exploits ResDense encoding blocks and a Dice-Focal combined loss. Our method's performance was evaluated quantitatively, using the Dice Coefficient (DICE), Precision (PR), Recall (RE), and Matthew's Correlation Coefficient (MCC), and qualitatively, achieving more detailed predictions and maintaining a better degree of continuity. Our method's optimal performance was observed using anisotropic patches within a sliding window approach, and we addressed the problem of class imbalance with an appropriately weighted loss function.