Application of topological deep learning in staging liver fibrosis based on hepatobiliary phase MR images.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology, Affiliated Hospital of Qingdao University, Qingdao, China.
- Department of Radiology, Affiliated Hospital of Qingdao University, Qingdao, China. [email protected].
Abstract
Early-stage liver fibrosis is reversible; thus, accurate staging of liver fibrosis is of great importance.. Liver biopsy is the gold standard for the diagnosis of liver fibrosis; however, its invasive nature limits its clinical application. Accurate noninvasive staging of liver fibrosis remains challenging. Thus, there is an urgent need for a precise and reliable noninvasive diagnostic approach. A topological deep learning model combining topological data analysis and EfficientViT framework (TopoEfficientViT) was developed to effectively capture global contextual relationships in hepatobiliary phase MR images for assessment of staging liver fibrosis. Our retrospective study included patients with chronic liver diseases from June 2018 to July 2025, which was approved by our Institutional Review Board. Our data from a hospital site (Site A) in our institution were divided into a training set (n = 275) and a validation set (n = 75). The test set (n=65) was composed of data from a separate hospital site (Site B) in the same institution. A total of 415 patients (323 males, mean age 56.52±9.54 years; 92 females, mean age 58.86±9.68 years) were enrolled in the study. In the training set, the AUC and accuracy of TopoEfficientViT for staging liver fibrosis F1-2 vs. F3-4 were 0.942 and 0.906, respectively; for staging liver fibrosis F1-3 vs. F4 were 0.957 and 0.891, respectively. In the validation set, the AUC and accuracy for staging liver fibrosis F1-2 vs. F3-4 were 0.849 and 0.76, respectively; for staging liver fibrosis F1-3 vs. F4 were 0.861and 0.813, respectively. In the test set, the AUC and accuracy for staging liver fibrosis F1-2 vs. F3-4 were 0.829 and 0.769, respectively; for staging liver fibrosis F1-3 vs. F4 were 0.791 and 0.785, respectively. The TopoEfficientViT model had high predictive performance for staging liver fibrosis by integrating multi-scale persistent homology topology with convolutional networks.