Fully Automatic Artificial Intelligence Liver Anatomy Segmentation in the Management of Colorectal Liver Metastases.

INTRODUCTION

Artificial intelligence is gaining increasing interest in medical image segmentation, including liver cancer. However, the literature lacks model implementation in the setting of colorectal liver metastases for treatment planning.

MATERIALS AND METHODS

We collected the portal phase abdominal CT scan images from the Nice University Hospital hepatobiliary oncologic multidisciplinary discussion of 80 patients with colorectal liver metastases, before treatment. Data from 70 patients was exploited to train and test the nnU-Net model to automatically perform parenchyma, portal vein, hepatic veins, cava vein, and colorectal liver metastases segmentation. Data from the remaining 10 patients was used for external validation.

RESULTS

The Dice score for parenchyma segmentation was 0,964 and 0,955 in the test and validation dataset, respectively. For portal vein segmentation, a centerline Dice (clDice) of 0,758 and 0,736 was highlighted, while for hepatic veins it resulted to be 0,758 and 0,577. Cava vein segmentation showed a clDice of 0,805 and 0,734. Concerning colorectal liver metastases, the Dice score was 0,693 and 0,61.

CONCLUSION

The nnU-Net showed promising segmentation results, especially for liver parenchyma. Its task could be useful to help physicians decide which is the best treatment strategy based on individual anatomical characteristics and disease extension. Training the model on a larger dataset with the same characteristics could help improve segmentation performances.