Patient-specific prediction of 3D ablation zones via oncological feature-conditioned deep generative modeling: An in...

Accurate prediction of the ablation zone is essential for effective radiofrequency ablation (RFA) therapy. However, current clinical guidelines often rely on oversimplified ellipsoidal assumptions based on homogeneous tissue properties, failing to reflect the patient-specific variability in cancer morphology. This study proposes a deep generative modeling framework conditioned on MRI-derived oncological features and electrode placement information to generate in silico 3D ablation zone predictions.

Our model utilizes a U-Net-based generator within a conditional DCGAN framework to capture irregular, non-ellipsoidal thermal patterns arising from tissue-dependent thermal and electrical heterogeneity. Trained on synthetic ablation data generated from validated numerical simulations incorporating MRI-derived cancer morphology, it implicitly learns the influence of biophysical heterogeneity. The model accuracy was evaluated using intersection-over-union (IoU), precision, sensitivity, and specificity metrics and tested on unseen patient data.

The model achieved high accuracy with an IoU of 90.92±3.54% and generated predictions in under 0.5 s, suggesting potential for integration into clinical decision-support workflows. It also demonstrated strong generalization to unseen patient data.

This study demonstrates the feasibility of deep generative modeling conditioned on MRI-derived oncological features for generating in silico 3D ablation zones. By leveraging biophysically validated simulation data, this model enables the virtual simulation and comparison of treatment outcomes from various electrode configurations and cancer morphologies, thereby refining the therapeutic strategy for each individual.