Document resource
Introduction Recurrence after thermal ablation in early-stage hepatocellular carcinoma (HCC) is high with a lack of clinical predictors. This study used pretreatment multiphase MRI-derived radiomic features to predict recurrence after ablation. Targeted RNA sequencing was used to determine drivers of radiomic signatures.Methods 156 HCC patients undergoing ablation from the SORAMIC ( NCT01126645) and Lausanne (NCT02859753) prospective phase 2 randomised control trials were reviewed. Radiomic features from pretreatment T1-weighted arterial, portal venous and delayed phase lesions were extracted. Radiomic models predicting 12-month recurrence for each phase were trained in the SORAMIC cohort using eight supervised machine learning models, including the state-of-the-art generative foundation model, TabPFN, and thirteen feature selection techniques. The best performing radiomic models were combined and integrated with clinical variables using ensemble learning. Patients were stratified using optimal threshold tuning. Model performance was externally validated in the independent Lausanne cohort. RNA sequencing was performed for radiomic risk-groups.Results The ensemble machine learning-based arterial phase radiomic model outperformed all clinical benchmarks and TabPFN in predicting recurrence in training (AUC 0.90, 95% CI 0.80-0.98) and external validation (AUC 0.74, 95% CI 0.61-0.87) cohorts. Arterial-portal venous and integrated radiomic-clinical models had comparable performance to the standalone arterial phase radiomic model. The arterial phase radiomic model stratified high-risk group had a significantly shorter median time-to-recurrence compared to the low-risk group in training (4.2 months, 95% CI 1.9-9.9 vs. 29.0 months, 95% CI 15.8-34.4; p<0.001) and external validation (4.5 months, 95% CI 1.2-9.4 vs. 12.5 months, 95% CI 9.6-22.9; p=0.007) (figure 1). Radiomic risk group was the only significant predictor of recurrence in multivariable Cox regression. Telomerase reverse transcriptase (TERT) expression was significantly enriched in the high-risk radiomic group.Conclusions Arterial phase radiomic-based models can predict recurrence after curative HCC ablation. Pretreatment MRI can identify patients at high-risk of recurrence for precision surveillance and adjuvant strategies.Abstract O13 Figure 1