Personalized modeling for in vivo multimodal thermal therapy planning based on preoperative MRI

Abstract Background Accurate treatment dosage planning depends on the personalized thermal response in vivo. Purpose This study proposes a personalized model by incorporating personalized thermal conductivity derived from the preoperative magnetic resonance imaging (MRI) characteristics of targeted tumors. Methods Ten female New Zealand rabbits with subcutaneous VX2 tumors underwent multimodal thermal therapy. An individual coefficient () was determined from reconstructed temperature mapping to estimate the effective thermal conductivity. Preoperative T2‐weighted (T2W) and T1‐weighted (T1W) MRI images were acquired, with defined as the ratio of mean signal intensity of the T2W images to that of the T1W images. The ratio of the value between the tumor and the adjacent muscularis propria (represented as ) was used to establish a quantified relationship with the individual coefficient (). The fitness using linear function and several common nonlinear functions were compared in the leave one out cross validation (LOOCV) to assess the precision and degree of overfitting. Additionally, a sensitivity analysis on the linear fitting to variations on MRI characteristics was conducted. Furthermore, the temperature distributions in tumors with varied parameters and freezing temperatures were predicted for preoperative planning. Results Effective thermal conductivity was fitted as a weighted combination of water (weighted by an individual coefficient , ranging from 0.5 to 0.75), and a baseline value of 0.16 W/m/K (weighted by ). The root mean square errors (RMSE) between reconstructed and experimental temperatures were 3.40 2.32 and 0.80 0.53 in transient and steady states, respectively. Quantified correlation between and the individual coefficient was explored and validated. Compared to the nonlinear correlations studied, no significant difference in residuals was observed, while the fitness using quadratic or cubic functions exhibited signs of overfitting with significantly increased AIC/BIC. The linear function was selected for prediction and sensitivity analysis. Using the predicted effective thermal conductivity, the RMSE were 3.79 2.16 and 1.24 0.46 in the transient and steady states, respectively, significantly lower than those from conventional model ( p < 0.01 and p < 0.05, respectively). The model remained robust to 10% perturbation error on preoperative MRI characteristics. In large tumors with low individual coefficient , personalized model reduced overestimation on ‐20 isothermal region by 15.11 mm compared to the conventional model. Conclusions Preoperative MRI provided a quantified and user‐friendly tool for improving accuracy of personalized model on thermal dosage planning.