Performance analysis of liver segmentation using nn-UNet TotalSegmentator: Focus on atypical livers, pathologies, and...

This study evaluates the accuracy of the nn-UNet TotalSegmentator (TS) by Wasserthal et al. (2023) in segmenting atypical livers with pathologies and variants in CT scans.

CT scans were retrospectively collected from our RIS and divided into two cohorts: a reference group (67 healthy livers) and a study group (55 scans across eleven pathology and variant subgroups). TS performed automatic segmentation for all groups. For reference, the images were then manually segmented, with corrections reviewed by two radiologists. Accuracy was assessed using Dice similarity score, Hausdorff distance (HD), mean surface distance (MSD), volume difference, and clinical ratings.

Automatic segmentation underestimated liver volume by a mean of 48.11  ml (3.1%) in the reference group and overestimated it in 84% of study group cases by 79.09 ml (4%). The average Dice score was 0.980 ± 0.007 for the reference group and 0.933 ± 0.113 for the study group. Hepatomegaly achieved the highest score (0.979 ± 0.006), Polycystic liver disease (PLD) the lowest (0.656 ± 0.230). Cirrhosis with Ascites, Beavertail, and PLD had significantly lower Dice scores than the reference group. Clinical ratings were often lower than Dice scores suggested, especially in Beavertail, Cirrhosis with Ascites, Ablation defects, Metastases, and Hemihepatectomy.

TS performs excellently on healthy and well on most pathological livers. Despite high Dice scores in many pathological cases, clinical ratings reveal limitations. Clinical evaluation remains essential. Inclusion of PLD and Beavertail cases in training data may reduce bias and improve performance.