Transfer-Learning for Cross-Weighted MRI Liver Cirrhosis Segmentation Using CirrMRI600+

Abstract

Accurate and fully automated liver segmentation has the potential to greatly help in the improvement of disease detection and the personalization of treatments. Liver cirrhosis is the final stage of liver disease and is characterised by fibrosis and the remodeling of the liver, which increases the mortality rate. Magnetic resonance imaging (MRI), particularly T1- and T2-weighted sequences, provides a powerful non-invasive tool for assessing liver structure and pathology. However, manual delineation of cirrhotic liver regions is both time-consuming and prone to human error, with performance heavily dependent on the pathologist’s expertise. Reliable segmentation is further complicated by pronounced morphological distortions and heterogeneous signal characteristics associated with cirrhosis. In this study, we address these challenges using a fully convolutional ResNet-UNet framework that employs cross-weighted transfer learning, retrained on T1-weighted MRI data to capture general anatomical representations and fine-tuned on T2-weighted images to adapt to contrast and pathological variability, using the CirrMRI600+ dataset. The proposed model achieved the highest Dice coefficient of 0.89 and mIoU scores of 0.81 on the testing set, surpassing the performance of training a single ResNet-UNet on T1 and T2 separately, without transfer learning.

 This cross-weighted training paradigm demonstrates the effectiveness of contrast-domain pretraining for robust and efficient liver segmentation, paving the way for multimodal MRI integration in hepatic disease assessment and advancing progress toward automated cirrhosis staging and personalised therapeutic planning.