A digital twin for microwave liver treatment replanning

Technical Summary: A Digital Twin for Microwave Liver Treatment Re-planning

Problem Statement
Microwave Ablation (MWA) is a leading minimally invasive treatment for liver tumors, offering advantages over other thermal modalities such as shorter treatment times and reduced susceptibility to the "heat-sink" effect caused by blood vessels. However, successful treatment relies heavily on precise preoperative planning and accurate intraoperative antenna placement. In clinical practice, physiological factors such as breathing motion, patient repositioning, and tissue deformation make it practically impossible to position the antenna exactly as planned.

When misplacement occurs, adhering strictly to the original plan often results in incomplete tumor ablation and potential recurrence. Conversely, repositioning the antenna via reinsertion is time-consuming and carries significant risks, including tumor seeding (the spread of cancer cells along the needle track). Current clinical planning relies on manufacturer-provided geometrical models that ignore patient-specific tissue properties, often overestimating ablation zones. There is a critical need for a patient-specific simulation tool capable of intraoperative re-planning to correct for misplacement without requiring antenna reinsertion.

Methodology
The authors propose a "digital twin" framework comprising two main components: a high-fidelity MWA simulator and an optimization algorithm for re-planning.

1. Computational Model:

   • Physics: The system solves coupled microwave propagation and heat transfer equations. Heat transfer is modeled using the Pennes Bioheat equation, while microwave propagation is governed by frequency-domain Maxwell equations.
   • Patient Specificity: The model incorporates patient-specific dielectric and thermal properties derived from preoperative CT and MRI images. It accounts for temperature-dependent tissue changes, including coagulation, water evaporation, and the shutdown of blood perfusion ($\omega_b$).
   • Numerical Implementation: A Finite Element Method (FEM) scheme using the FEniCSx library is employed. To balance accuracy and computational cost, the authors utilize a weakly coupled approach where the microwave propagation is solved once, and the external heat source ($Q_{ext}$) is approximated as a temperature-dependent function. This avoids the need for re-solving Maxwell's equations at every thermal iteration.
   • Antenna Characterization: Since internal antenna geometries are proprietary, the model estimates antenna characteristics (e.g., choke dimensions) by optimizing the S11 reflection coefficient against known external geometries.
   • Validation: The model was validated against in vivo swine experiments involving three ablation scenarios: a baseline far from large vessels, one near the hepatic vein, and one near both hepatic and portal veins.

2. Re-planning Optimizer:

   • Objective: To correct antenna misplacement by adjusting three parameters: power input ($P$), ablation duration ($T$), and antenna translation along its axis ($\vec{t}$).
   • Algorithm: A Genetic Algorithm (GA) was selected due to the non-linear nature of the problem and the lack of gradient information. The objective function minimizes a weighted sum of the inverse Dice similarity coefficient and the 95th percentile Hausdorff distance ($dH_{95}$) between the simulated ablation zone and the target zone.
   • Constraints: The optimization operates within realistic clinical bounds: power (20–150 W), duration (up to 15 min), and translation ($\pm$5 cm).

Key Results

• Simulation Accuracy: The simulator demonstrated high accuracy in predicting ablation zones compared to ground truth from in vivo swine data. It achieved Dice similarity coefficients of 0.82, 0.81, and 0.79 for the baseline, hepatic vein, and portal/hepatic vein scenarios, respectively. These results significantly outperformed the manufacturer's geometrical models, which yielded Dice scores of 0.62, 0.66, and 0.58. The simulator also showed robustness against the heat-sink effect, maintaining low Hausdorff distances ($dH_{95} \approx 2$ mm) even near large vessels.
• Re-planning Performance: In a re-planning scenario where the antenna was artificially shifted by 5 mm, the optimizer successfully recovered parameters to match the target ablation zone. The re-planned simulations achieved Dice scores of 0.83, 0.83, and 0.80, representing average improvements of 20.3%, 40.7%, and 48.1% over the initial (misplaced) ablation results. The optimizer recovered power and translation parameters with low relative errors (e.g., <2.3% for power in baseline cases), even in complex vascular environments.

Significance and Claims
The paper claims to be the first work to address intraoperative re-planning of thermal ablation therapy. Its primary significance lies in demonstrating that complete tumor ablation can be achieved without antenna reinsertion by optimizing specific ablation parameters (power, time, and axial position).

The authors position this work as a proof-of-concept for a clinical decision-support tool. While acknowledging that the current computational time (50–95 minutes) precludes real-time use without further acceleration (e.g., via GPUs or surrogate models), the study validates the feasibility of using patient-specific digital twins to correct for inevitable intraoperative deviations. The work suggests that such a tool could prevent incomplete ablations and avoid the risks associated with tumor seeding, particularly in complex cases involving large vessels where standard planning models fail.

Limitations
The authors note that the current approach requires a baseline ablation to calibrate the blood perfusion parameter ($\omega_0$), which is not feasible in a clinical setting without prior intervention. They suggest that future work could integrate methods to estimate perfusion directly from CT images. Additionally, the current model does not account for heterogeneous tumor tissue, and the re-planning tool has not yet been tested on actual tumor models.