Background: This study focuses on an AI model for automatic segmentation of liver territories in living liver donors based on hepatic venous territories.
Methods: A total of 95 donor CTs were used for building an autosegmentation model. The segmentation data was generated by professional biomedical visualization artist and the liver was divided into venous territories. Right hepatic venous territories were divided into right superior and inferior hepatic veins. Middle hepatic vein territories were divided into V5 and V8 territoreis. When it showed definite borders, V4a and V4b territories were divided. Utilizing SwinUNETR v2 and trained on 95 CT cases, the model was evaluated with the Dice Similarity Coefficient (DSC).
Results: The approach yielded promising DSC scores across various regions, with an average of 81.0%. Regarding portal vein territories divided by four, the average DSC score was 87±2.6. Regarding hepatic venous territories divided by three, the average DSC score was 85.5±4.3. The DSC scores for V5 and V8 were 75.7±6.3 and 67.2±8.6, respectively.
Conclusion: The autosegmentation AI model showed promising results. Increasing the dataset size is expected to enhance performance, making this method potentially valuable for surgical planning and liver transplantation.