Simulation of Magnetic Resonance guided Laser Interstitial Thermal Therapy Temperature Maps through Time-series based Deep Learning Methods: Usage of ConvLSTM

Saba Sadatamin, Paola Driza, Gemma Postill, Steven Robbins, Richard Tyc, Rahul Krishnan, Lueder Alexander Kahrs, Adam C. Waspe, James M. Drake

Keywords: magnetic resonance-guided laser interstitial thermal therapy (MRgLITT), deep learning, convolutional long short term memory (ConvLSTM), treatment monitoring

Abstract: Magnetic resonance-guided laser interstitial thermal therapy (MRgLITT) is a minimally invasive therapy that leverages thermal ablation to treat drug-resistant focal epilepsy. Patient-specific heat sinks, such as blood vessels, complicate the planning of MRgLITT as it creates patient-level variability in how heat from the laser propagates, thus potentially undermining treatment efficacy. To simulate the MRgLITT temperature maps, we developed a deep learning framework which can predict the resulting spatio-temporal temperature maps of the monitoring system. We used a convolutional long short-term memory (ConvLSTM) network and evaluated the outcome using both quantitative vs. impact evaluation metrics. In impact evaluation, we binarized temperature images and pixels exceeding a temperature of 39$^{\circ}$C are identified as potentially indicating cell death. We then use a dice score and sensitivity metrics to evaluate the overlap between predicted and ground truth thermal dose margin. We demonstrate strong performance of our ConvLSTM framework, with a structural similarity index metric of 0.88, dice score of 0.85 and sensitivity of 0.77 which shows that predicted heat propagation was highly similar to the ground truth. Our findings can be used by neurosurgeons to improve the delivery of MRgLITT.

Submission Number: 146