A fully automated hybrid approach to assessing liver fibrosis and necroinflammation on conventional MRI: A multi-center cohort Study

Abstract

Abstract Background & Aims: To develop and validate the CoRC model at conventional MRI for diagnosing clinically significant liver fibrosis (≥ F2) and necroinflammation (≥ G2). Materials and methods: This retrospective cohort study recruited 537 patients with biopsy-proven liver fibrosis and necroinflammation at center 1 from May 2015 to Aug 2020 and center 2 between Jan 2011 and Jan 2021. 394 patients were randomly allocated into training (n = 276) and internal test (n = 118) cohorts at center 1.. Automated entire liver segmentation used ResUNet-based Human-in-the-Loop approach. Radiomics features were extracted from the mask on fat-suppressed T2-weighted and delayed enhanced T1-weighted images separately. Radiomics signatures were generated using logistic regression. Radiomics-scores and optimal clinical biomarkers as independent risk factors were integrated into the CoRC models in the training cohort with multivariate logistic regression. Models were tested in independent temporal test cohort at center 1 (n = 96) and an external test cohort from center 2 (n = 47). Diagnostic performance was evaluated by area under the curve, calibrations and decision curve analysis. Results: In the internal, temporal, and external test cohorts, CoRC model 1 yielded AUCs of 0.79, 0.82 , and 0.83 for ≥ F2, meanwhile, CoRC model 2 showed AUCs of 0.86, 0.79, and 0.89 for ≥ G2. ,We compared CoRC models with transient elastography-based liver stiffness measurement (TE-LSM) subgrouply (AUC 0.78 vs.0.79, P = 0.86/0.82 vs. 0.73, P = 0.14 for ≥ F2, whereas 0.88 vs. 0.81, P = 0.16/0.79 vs. 0.74, P = 0.49 for ≥ G2). Conclusions CoRC models exhibited promising diagnostic performances for ≥ F2 and ≥ G2, which could be a potential alternative when TE-LSM is unavailable.