A PSC-tailored deep learning model for liver segmentation on fat-saturated T2-weighted MR: Robustness to hepatic dysmorphia and multicentre generalisability.

In this study, we aimed to identify an effective and accurate liver segmentation strategy for fat-saturated T2-weighted MRI in PSC patients, enabling multicentre validation of GLRM-RunEntropy, a radiomics feature which showed prognostic value. We evaluated three segmentation tools (Total Segmentator, Philips MR Liver Health, and an in-house trained nnU-Net model) across 183 PSC patients from internal and external cohorts. Ten segmentation algorithms derived from these tools were quantitatively assessed using DSC and HD against reference liver contours. The nnU-Net model trained on PSC-specific T2w-SPIR images (algorithm #10) outperformed all others, achieving a mean DSC of 95 % and mean HD of 25.7 mm. Total Segmentator (algorithm #1) showed strong DSC (90 %) but was more affected by variations in liver morphology. Qualitative evaluation by expert radiologists confirmed the superior inclusion of liver parenchyma and exclusion of non-hepatic tissues by the nnU-Net model, together with a lower sensitivity to liver stiffness and robustness to cohort variability. Our findings highlight the importance of domain-specific training for robust segmentation in the setting of chronic liver disease imaging. The validated nnU-Net segmentation model will be made available to support large-scale multicentre validation of GLRM-RunEntropy as a prognostic biomarker in PSC.