Subscribe to RSS
DOI: 10.1055/s-0045-1802893
Performance Analysis of Liver Segmentation Using nn-UNet TotalSegmentator: Focus on Atypical Livers, Pathologies, and Variants
Authors
Zielsetzung This study evaluates the accuracy of the nn-UNet TotalSegmentator for segmenting atypical livers with pathologies and variants in CT scans.
Material und Methoden We gathered CT scans retrospectively and grouped them into two cohorts: one with healthy livers (72 scans) and another with 55 scans across eleven pathology and variant subgroups including beaver tail liver, hemihepatectomies, ablation defects, hepatomegaly, steatosis hepatis, cirrhosis, cirrhosis with ascites, polycystic liver disease (PLD), metastasized liver, and pediatric scans. The TotalSegmentator performed the automatic segmentation across all groups. As a reference, the images were then manually segmented with corrections reviewed by two radiologists. Six Metrics were used to assess the accuracy of the TotalSegmentator: Dice score, Hausdorff distance, mean surface distance, total volume difference, and a clinical radiologist rating.
Ergebnisse For manual segmentation, the average mean volume for the liver cohort was 1548 ml. Automatic segmentation typically overestimated the volume by 3.09% in healthy livers and underestimated it by 13.13% in pathological subgroups. The total average Dice score for the reference group with healthy livers is 0.98±0.007 whereas the total average Dice score for the study group is 0.93±0.113, which indicates a significant difference between the groups. Notably, the hepatomegaly subgroup showed the highest Dice score (0.98±0.006), while PLD had the lowest (0.66±0.23). Clinical ratings are frequently lower than Dice scores suggest. Although Dice scores were above 0.9, clinical ratings were often insufficient.
Schlussfolgerungen The automatic segmentation with the nn-Unet TotalSegmentator excels in healthy liver segmentation. For pathological liver CT scans, mainly high Dice scores were calculated. Clinical assessments, on the other hand, indicate that quantitative measures are not enough to assess a segmentation tool’s usefulness in clinical practice.
Publication History
Article published online:
25 March 2025
© 2025. Thieme. All rights reserved.
Georg Thieme Verlag KG
Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany