<p>Thermal ablation therapy has emerged as a promising strategy for treating tumors that are difficult to resect surgically, such as deep-seated glioblastoma. Recent advances in imaging modalities, particularly photoacoustic and ultrasound imaging, have demonstrated the feasibility of direct necrosis monitoring, offering more accurate assessments than traditional temperature-based methods. Building on these developments, the integration of real-time necrosis feedback (NFB) into ablation control has been introduced. However, several existing NFB control studies neglected the influence of residual heat after ablation is terminated. Furthermore, NFB itself lacks temperature monitoring capabilities, raising key questions regarding how residual heat affects ablation outcomes and how it should be effectively managed. Model predictive control (MPC) offers a potential solution for managing residual heat, but its effectiveness depends on the accurate identification of patient-specific thermal parameters. To address these challenges, we propose a Direct Necrosis-Monitoring-based Adaptive Model Predictive Control (DNaMPC) framework. This method leverages real-time NFB while accounting for residual heat and adaptively identifies patient-specific parameters using a hierarchical Extended Kalman Filter (EKF) architecture. The approach employs a novel parameter absorption strategy, where a single highly observable parameter (thermal damage threshold, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\tau$</EquationSource> <EquationSource Format="MATHML"><math> <mi>τ</mi> </math></EquationSource> </InlineEquation>) is adaptively estimated to compensate for uncertainties in both thermal diffusivity (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\alpha$</EquationSource> <EquationSource Format="MATHML"><math> <mi>α</mi> </math></EquationSource> </InlineEquation>) and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\tau$</EquationSource> <EquationSource Format="MATHML"><math> <mi>τ</mi> </math></EquationSource> </InlineEquation> itself, circumventing the identifiability challenges of simultaneous multi-parameter estimation. In simulations based on a one-dimensional (1D) finite difference model of brain tissue, which serves as a thermodynamic worst-case scenario for residual heat management, DNaMPC was compared against proportional (P) control and Recursive Least Squares (RLS)-based adaptive MPC (RLS-MPC). DNaMPC eliminated post-shutdown excessive ablation observed with P-control and outperformed RLS-MPC in terms of targeting accuracy and input energy economy. Further evaluation under model uncertainties demonstrated the framework’s robustness, achieving the target ablation ratio of 1 despite up to <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$\pm 8\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>±</mo> <mn>8</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> mismatch in thermal diffusivity (<InlineEquation ID="IEq5"> <EquationSource Format="TEX">$\alpha$</EquationSource> <EquationSource Format="MATHML"><math> <mi>α</mi> </math></EquationSource> </InlineEquation>) and <InlineEquation ID="IEq6"> <EquationSource Format="TEX">$-30\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>-</mo> <mn>30</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> to <InlineEquation ID="IEq7"> <EquationSource Format="TEX">$+270\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>+</mo> <mn>270</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> mismatch in the thermal damage threshold (<InlineEquation ID="IEq8"> <EquationSource Format="TEX">$\tau$</EquationSource> <EquationSource Format="MATHML"><math> <mi>τ</mi> </math></EquationSource> </InlineEquation>). The adaptive control demonstrated superior performance in terms of both median accuracy and reduced variability compared to non-adaptive approaches across the expected range of patient parameter variations, maintaining robustness against sensing noise levels up to 20%. These results demonstrate that DNaMPC can effectively utilize residual heat, suppress sensing errors, and adapt to patient-specific variability during the procedure, thereby laying the foundation for precise, individualized thermal ablation therapies.</p>

Direct-Necrosis-Monitoring-Based Adaptive Model Predictive Control for Ablation Therapy Including Patient-Specific Residual Heat Management

期刊论文
发表日期：

2026年3月26日

查看全文

Ryo Murakami,
Satoshi Mori,
Haichong K. Zhang

摘要

Thermal ablation therapy has emerged as a promising strategy for treating tumors that are difficult to resect surgically, such as deep-seated glioblastoma. Recent advances in imaging modalities, particularly photoacoustic and ultrasound imaging, have demonstrated the feasibility of direct necrosis monitoring, offering more accurate assessments than traditional temperature-based methods. Building on these developments, the integration of real-time necrosis feedback (NFB) into ablation control has been introduced. However, several existing NFB control studies neglected the influence of residual heat after ablation is terminated. Furthermore, NFB itself lacks temperature monitoring capabilities, raising key questions regarding how residual heat affects ablation outcomes and how it should be effectively managed. Model predictive control (MPC) offers a potential solution for managing residual heat, but its effectiveness depends on the accurate identification of patient-specific thermal parameters. To address these challenges, we propose a Direct Necrosis-Monitoring-based Adaptive Model Predictive Control (DNaMPC) framework. This method leverages real-time NFB while accounting for residual heat and adaptively identifies patient-specific parameters using a hierarchical Extended Kalman Filter (EKF) architecture. The approach employs a novel parameter absorption strategy, where a single highly observable parameter (thermal damage threshold, $\tau$ $τ$ ) is adaptively estimated to compensate for uncertainties in both thermal diffusivity ( $\alpha$ $α$ ) and $\tau$ $τ$ itself, circumventing the identifiability challenges of simultaneous multi-parameter estimation. In simulations based on a one-dimensional (1D) finite difference model of brain tissue, which serves as a thermodynamic worst-case scenario for residual heat management, DNaMPC was compared against proportional (P) control and Recursive Least Squares (RLS)-based adaptive MPC (RLS-MPC). DNaMPC eliminated post-shutdown excessive ablation observed with P-control and outperformed RLS-MPC in terms of targeting accuracy and input energy economy. Further evaluation under model uncertainties demonstrated the framework’s robustness, achieving the target ablation ratio of 1 despite up to $\pm 8\%$ $\pm 8 %$ mismatch in thermal diffusivity ( $\alpha$ $α$ ) and $-30\%$ $- 30 %$ to $+270\%$ $+ 270 %$ mismatch in the thermal damage threshold ( $\tau$ $τ$ ). The adaptive control demonstrated superior performance in terms of both median accuracy and reduced variability compared to non-adaptive approaches across the expected range of patient parameter variations, maintaining robustness against sensing noise levels up to 20%. These results demonstrate that DNaMPC can effectively utilize residual heat, suppress sensing errors, and adapt to patient-specific variability during the procedure, thereby laying the foundation for precise, individualized thermal ablation therapies.