An MRI-Based Intratumoral and Peritumoral 2.5D Deep Learning Model for Predicting P53-Mutated Hepatocellular Carcinoma: A Two-Center Study.

Abstract

IntroductionHepatocellular carcinoma (HCC) is the most common primary liver cancer with high mortality. TP53 is the most frequently mutated tumor-suppressor gene in HCC, and P53 mutations are associated with poor differentiation, vascular invasion, and lower survival. Preoperative noninvasive detection of P53-mutated HCC is critical for treatment planning and prognostic assessment. This study aimed to develop and validate a deep learning (DL) model for predicting P53 mutation status in HCC.MethodsIn this retrospective two-center study, 373 patients with pathologically confirmed HCC were enrolled from two institutions (Center 1: n=320; Center 2: n=53). The Center 1 cohort was randomly partitioned into a training cohort (n=256) and an internal test cohort (n=64). Patients from Center 2 constituted an independent external test cohort (n=53). We employed a 2.5D ResNet34 architecture to process three-channel dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Separate deep learning models were constructed for intratumoral and peritumoral regions using both single-phase and multiphase (arterial, portal venous, and delayed) fusion sequences. Significant clinical variables were selected to build a Clinical model, which was subsequently integrated with the optimal deep learning model to create a Combined model. Model performance was assessed by the area under the curve (AUC), calibration curves, and decision curve analysis (DCA).ResultsMultiphase fusion models consistently outperformed their single-phase models. The peritumoral (5mm) fusion model (IntraPeri5mm_Fusion) outperformed the intratumoral fusion model, achieving AUCs of 0.881 (95% CI: 0.836-0.926), 0.772 (95% CI: 0.651-0.893), and 0.757 (95% CI: 0.624-0.891) on the training cohort, internal test cohort, and external test cohort, respectively. The Combined model yielded further incremental improvement, with AUCs reaching 0.886 (95% CI: 0.842-0.930), 0.805 (95% CI: 0.691-0.918), and 0.793 (95% CI: 0.668-0.918) across the respective cohorts.ConclusionsThe proposed 2.5D DL model based on DCE-MRI effectively predicts P53-mutated HCC. Integrating peritumoral features with clinical variables yields the best performance, offering a promising noninvasive tool for prognostic stratification and individualized therapeutic planning.