Patient-specific biophysical AI-powered modeling for predicting RFA volumes in HCC.

Published in European radiology experimental. Abstract: To clinically evaluate a fully automated, artificial intelligence-powered, patient-specific biophysical model for predicting radiofrequency ablation (RFA) volumes in hepatocellular carcinoma (HCC) and to compare its performance with manufacturer-provided ablation estimates. This retrospective study included 96 RFA procedures in 65 HCC patients treated using a single ablation with a cluster needle. Tumors, needle position, and postablation zone were manually segmented on periprocedural computed tomography. Liver parenchyma and vascular structures were automatically extracted using deep learning. A graphics processing unit (GPU)-accelerated biophysical model simulated heat transfer, vascular heat-sink effects, and tissue necrosis using patient-specific anatomy and procedural data. Correlations between ablation volume manually segmented and clinical or anatomical parameters (ablation time, ablation tip temperature, tumor-capsule distance, vessel proximity, type, and diameter) were assessed. Model-predicted volumes were compared with ground truth and manufacturer estimates using relative volume difference (RVD) and volume similarity. Final ablation volumes showed wide variability (7.9-78.8 cm³) and were not correlated with any individual clinical or anatomical parameter (Pearson r - 0.02 to 0.29). Across 94 evaluable cases, simulated and ground-truth volumes were similar (p = 0.165), while manufacturer estimates significantly overestimated ablation volume (p < 0.001). The model a...