Noninvasive Tumor Profiling: Quantitative Contrast-Enhanced MRI Markers Predict PD-L1 and CTNNB1 Status in Hepatocellular Carcinoma

Abstract

Background: Systemic targeted therapy (STT) is the standard of care for patients presenting with advanced-stage hepatocellular carcinoma (HCC); however, tumor response rates are limited, mainly owing to HCC biomolecular and pathological heterogeneity. Therefore, novel markers for noninvasive molecular profiling are needed. Purpose: To determine whether advanced image analysis and machine learning on routinely acquired MRI can predict HCC molecular profiles, thereby allowing biomarker-guided treatment allocations. Methods: This single-center retrospective study included treatment-naïve patients with HCC who underwent tumor resection or liver transplantation between 09/2006 and 02/2022. Multiparametric contrast-enhanced MRI data were obtained, and quantitative and qualitative imaging markers were extracted from lesion and liver segmentations. Pathology analysis of the resected samples was performed via immunohistochemistry to assess p53 loss of function; CTNNB1 activation; FoXM1 activation; and PD-L1, pAKT, pSMAD2/3 and SOAT1 expression. For each molecular profile outcome, a multivariable logistic regression model was built separately using quantitative imaging, qualitative imaging, or clinical data. The area under the receiver operating characteristic curve (AUC) was used to evaluate model discriminatory performance, and DeLong tests were performed to compare AUCs across models trained on the different data. Results: Seventy-five patients with T1-weighted, contrast-enhanced, dynamic MRI (mean age, 65.7 years ± 9.43 [SD]; 60 males) were included. ROC curve analysis demonstrated the good discriminatory performance of logistic regression models trained on quantitative imaging data for PD-L1, p53, and CTNNB1, with AUC values of 0.85 (95% CI: 0.74-0.96), 0.79 (95% CI: 0.66-0.93), and 0.7 (95% CI: 0.46-0.93), respectively. Models trained on clinical and qualitative imaging data yielded lower AUC values across profiles, of 0.36 (95% CI: 0.2-0.53) (P