Liver cirrhosis refers to a chronic liver disorder wherein the healthy hepatic tissue is replaced with fibrotic scar tissue. Liver cirrhosis impedes healthy liver functions in this process. For a complete evaluation of this chronic condition, precise and automated liver segmentation becomes essential. Apart from clinical assessment, accurate liver segmentation can also support longitudinal monitoring, biomarker detection, and early treatment planning. This study examines the risk of liver cirrhosis and introduces HepaLite-FAU, a lightweight U-net architecture featuring axial factorized convolutions and squeeze-and-excitation (SE) channel recalibration, for efficient and robust liver segmentation from MRI scans. The axial factorization approach dissects 3D/2D convolutional kernels into orthogonal 1D components. This acts as a fine line between computational efficiency and preserving precise contextual representation. The integration of SE ensures preservation of liver-specific feature channels. This approach also eliminates excessive features from the background. The proposed model has been evaluated on the CirrMRI600+ multi-sequence dataset. This dataset contains T1-weighted 3D MRI scans, T2-weighted 3D MRI scans, and T2-weighted 2D MRI scans. HepaLite-FAU has demonstrated consistent generalization capabilities for every scan type and proved to be strong with state-of-the-art performance with Dice scores of 95.73% on T1-3D, 92.35% on T2-3D, and 85.02% on T2-2D scans. The study reached correspondingly strong IoU, Precision, and Recall metrics as well. It also achieved robustness through stratified analysis across cirrhosis severity levels (mild, moderate, and severe). Overall, the results highlight that HepaLite-FAU provides a lightweight yet accurate solution for liver segmentation across multi-sequence MRI scans.
Jeem et al. (Fri,) studied this question.