Optimization of hepatic vasculature segmentation from contrast-enhanced MRI, exploring two 3D Unet modifications and various loss functions

Although MRI with hepatospecific contrast agents is a new standard diagnostic imaging for patients with liver cancer, there are no automated methods for detailed segmentation of the liver vasculature in cases with progressed tumors. This is due to the anatomical complexity, underlying disease and tunability of MRI image contrast, which challenge automatization. Here, we investigated the feasibility of liver vessel segmentation with three CNN architectures in combination with four different loss functions. In particular, a 3D Unet, a Vnet and its modification with an intra-layer dense block (DVnet) were evaluated. Dice-based loss, categorical cross entropy (CCE), weighed categorical cross entropy (WCCE) and focal loss (FL) were used as loss functions for training, the latter two to deal with the imbalanced class problem. A cohort of 90 patients (60 training, 10 validation and 20 testing) with progressed liver tumors were involved in this study, with manually annotated liver vasculature as a “gold standard”. Trained networks were evaluated by means of the Dice coefficient and centerline-based F1 score calculations. Models trained with balanced loss functions (FL,WCEE) performed the best for DVnet, while Vnet had the best performance for unbalanced loss functions. Vnet and DVnet architectures trained with an FL had the best overall segmentation accuracy (DC = 70%), while networks with a Dicebased loss had the lowest performance (max DC = 42%). In conclusion, the use of balanced loss functions, addition of an intra-layer dense block and drop-outs into the network architecture improved handling the unbalanced class problem in liver vessel segmentation.