Deep learning-based contouring of Couinaud segments on CT: Utility for volumetric analysis of future liver remnant.

Abstract

4113 Background: Hepatocellular carcinoma (HCC) is the most common primary liver tumor, and the liver is a frequent site of metastasis of other cancers. High tumor burden, proximity to hepatic vessels, and other comorbidities render only 30% of patients as candidates for curative treatment: transplantation, resection, or ablation. Surgical resection requires a 20 - 40% future liver remanent (FLR) to avoid post-operative complications. Delineation of the Couinaud segments is essential for volumetric analysis of FLR and targeted localization of tumors during pre-surgical treatment planning. Currently, manual annotation of the Couinaud segments in one CT volume can take two or more hours, which makes it cumbersome, and can benefit from automation. The purpose of this work is to develop an automated Couinaud segmentation tool for fast and accurate FLR estimation. Methods: Three CT datasets were used: 1) 161 patients from the public Medical Segmentation Decathlon (MSD) Hepatic Vessels dataset, 2) 43 patients with cirrhosis and metabolic diseases having ascites and splenomegaly imaged at the National Institutes of Health (NIH), and 3) 197 patients in the public TCIA Colorectal Liver Metastasis (CRLM) dataset. FLR annotation in the CRLM dataset was done by an expert radiologist. The Couinaud segments in the MSD and NIH datasets were manually annotated by two physicians using ITK-SNAP. The MSD and NIH datasets were used for training, while the CRLM dataset was reserved for testing. A 3D nnU-Net model was trained with default hyperparameters to outline the Couinaud segments. On the test dataset, the predicted Couinaud segments were overlaid on the FLR annotation, and metrics, such as Dice Similarity Coefficient (DSC), Hausdorff Distance (HD) error (in mm), and volume error (in cc) were calculated. The performance was compared to a previously described 3D U-Net model developed to quantify liver segmental volume ratio (LSVR) in patients with cirrhosis. Results: The 3D nnU-Net obtained a DSC of 0.99 ± 0.01 (IQR: 0.991, 0.998), HD error of 0.87 ± 1.83 mm (IQR: 0, 1.02), and volume error of 13.7 ± 28.1 cc (IQR: 3.4, 15.3). In contrast, the LSVR U-Net model attained a DSC of 0.97 ± 0.01 (IQR: 0.972, 0.984), HD error of 9.56 ± 3.61 mm (IQR: 7.14, 11.93), and volume error of 47.9 ± 51.8 cc (IQR: 24.5, 55.4). The 3D nnU-Net model achieved significantly different results for DSC (p < 0.001, large effect size 0.86), HD error (p < 0.001, large effect size 0.87), and volume error (p < 0.001, large effect size 0.91). Conclusions: The model showed acceptable generalizability to the external TCIA CRLM dataset. Future work may be directed towards accurate volumetric analysis on patients undergoing portal vein embolization to increase the FLR, and automatic tumor localization on specific Couinaud segments.