Preoperative Magnetic Resonance Imaging Based Predictive Modeling of Brain Tumor Laser Ablation

Abstract

Abstract INTRODUCTION Laser interstitial thermal therapy (LITT) yields precise lesions under magnetic resonance imaging (MRI) thermometry guidance and is revolutionizing brain tumor management. However, one limitation is the inability to preoperatively predict how tumors will respond to thermal energy. Our study assessed whether preoperative variables correlated with intraoperative LITT ablation dynamics were capable of predicting how ablations progressed. METHODS Pixels indicating irreversible damage were quantified as a function of time for 101 brain tumor patients treated with LITT. Ablation dynamics were determined using pixel counts modeled with first-order dynamics and related to independent preoperative variables using stepwise regression. The resulting models were evaluated for predictive accuracy using leave-one-out cross-validation analysis. RESULTS Total pixels ablated is larger for LITT performed at a high laser power and for a long duration and for lesions with a low T1 GAD signal. The rate of pixel increase is elevated with high laser power and for lesions with a low T2 signal and prior radiotherapy. Pixel counts are well modeled with first-order dynamics. Scale factor (C) is positively related to perfusion, laser power, and T2 signal and negatively related to T1 signal. Time constant (t) is positively related to T2 signal and perfusion. Tshift is negatively related to T2 signal and laser power. Pathology did not impact ablation dynamics, although the presence of radiation necrosis resulted in faster and larger ablations. A predictive model based on the independent variables accounts for 77% of the variance in ablation pixel counts. CONCLUSION Preoperative MRI features correlate with brain tumor LITT ablation dynamics. Predictive models based on these features may eventually be used to guide the planning and delivery of LITT.