Learning ablation volumes of laser interstitial thermal therapy (LiTT) for computer-assisted planning in minimally invasive neurosurgery
Project reference: DTP_SIE_06
First supervisor: Rachel Sparks
Second supervisor: Sebastien Ourselin
Start Date: October 2021
Project summary: The aim of this project is to develop novel artificial intelligence methods to accurately predict the treatment volume for laser interstitial thermal therapy (LiTT) applied to neurosurgery. LiTT allows for minimally invasive focal destruction of tissue around a laser catheter that is placed using keyhole stereotactic surgery. The current standard of care is for an expert neurosurgeon to determine the best placement of the laser, typically using a crude geometric estimation of a cylinder around the laser catheter. However, brain vascularity, proximity to bone and the ventricles results in asymmetric tissue heating which result in ablation volumes which either risk damaging critical structures (brain stem, optic and facial nerves) or insufficiently ablating targets of interest. We aim to develop deep learning models to learn the expected ablation volume using patient anatomy as input to the model. This will enable computer-assisted planning of LiTT trajectories which optimise ablating the treatment volume and minimise risk of damaging surrounding critical structures to provide a more accurate estimation of tissue damage.
Project description: Laser interstitial thermal therapy (LiTT) is a novel therapy which may provide a minimally invasive means of ablating structures within the brain. Thermal ablation is a lesioning technique that has been used in neurosurgery for many years with variable success. The main limitation to earlier methods were the unpredictable nature of thermal lesioning and the lack of real-time monitoring. LiTT is a laser catheter technique that allows for real-time monitoring of lesioning under MR thermography. However, a critical part to the process, both in terms of safety and efficacy, involves pre-operative planning of the laser trajectory as this ultimate determines ablation location and volume.
Current computer-assisted planning techniques assume crude geometric approximation of the ablation volume, typically as a 15mm diameter around the laser catheter. However, brain vascularity, proximity to bone and the ventricles result in asymmetric tissue heating. Developing computational models to more accurately estimate the heat transfer, and ultimately predict ablation volume would enable more accurate placement of laser catheter, could provide guidance on the heating parameters to use to obtain optimal ablation results. We propose to develop deep learning techniques to predict ablation volume and use these models to identify optimal laser catheter placement.
The student will develop tools to predict ablation volume for selective lesioning of the amygdalohippocampal complex using a multi-center retrospective dataset of over 100 LiTT procedures. These procedures have a clearly defined trajectory approach, limiting the variability of the ablative volume between patients. For a limited set of patients (20) we have corresponding MR thermography data over the course of the procedure with which to learn and validate the course of heating.
Second, prediction of heating using the trained network will be integrated into a computer-assisted planning software for the trajectory. The model will predict expected heating and will be used to aid surgeons in planning trajectories. A prospective trial taking place at the National Hospital for Neurology and Neurosurgery will be used to validate the accuracy of the predicted ablation volume.
Finally, if ablation prediction is accurate, we will investigate expanding this method to other ablative procedures including corpus callosotomy and hypothalamic hamartoma where the approach and anatomy are more variable and may require more complex models of heat dissipation and its relationship to anatomy.