Real-time patient-specific microwave ablation zone prediction via a unified bioheat solver and MRI-informed perturbation learning.

Microwave ablation is a crucial option for liver tumors, with success hinging on generating a suitably sized ablation zone for complete tumor eradication. Mathematical modeling supports ablation zone prediction and clinical decision-making, but its utility is limited by high computational cost and modeling complexity. We propose a physics-informed ablation modeling framework based on a homogenization-perturbation strategy that separates device-driven physics from patient-specific variability. A unified bioheat solver efficiently computes a semi-analytical temperature baseline induced solely by device-configurable parameters. A dual-branch network learns first-order corrections from pre-ablation MRI to align this baseline with individual anatomy and physiology. Ground-truth ablation zones delineated on post-ablation MRI provide high-fidelity image-based supervision. The framework combines physical interpretability with data-driven adaptability, enabling real-time (∼20 ms), patient-specific prediction and offering broad potential for clinical deployment and extension to other thermal therapies.