Artificial Intelligence acceleration of BNCT dose calculations

G. MARZIK; M.E. CAPOULAT; A.J. KREINER; D.M. MINSKY

Cracovia

Congreso; 20th International Congress on Neutron Capture Therapy, ICNCT 20; 2024

International Society for Neutron Capture Therapy

Trabajo ya aceptado (se adjunta mail de aceptación) - a presentar en junio 2024In the context of Boron Neutron Capture Therapy (BNCT), the impact of theneutron beam employed for patient irradiation extends beyond the neutroncapture reaction with the 10B isotope associated with tumor cells. Variousinteractions involving neutrons or secondary photons with other elementscould potentially affect healthy tissues. Hence, meticulous treatmentplanning is of paramount importance, ensuring that a given neutron beamconfiguration optimizes the probability of tumor control while minimizingadverse effects on healthy tissues. This optimization is achieved throughthe calculation of dose maps over tumor and healthy tissues.Traditionally, these dose maps are estimated using neutron transportsimulations based on Monte Carlo methods, which, while accurate, entail asubstantial computational cost and demand high computational power andlong simulation times for convergence and low statistical errors [1]. Thiscomplexity poses challenges for comprehensive studies on optimal treatmentconfigurations for individual patients and may impede the widespreadadoption of this therapy in medical centers aiming to treat multiplepatients daily.This study introduces a novel approach leveraging a neural network modeldesigned to expedite the convergence of Monte Carlo simulations. Theprimary objective is to shift the time-intensive aspect of simulations tothe training of the neural network, a process performed only a limitednumber of times. The proposed model is built upon a variant of the U-Netarchitecture, as illustrated in Figure 1.The input data comprises the CT scan of a patient, divided in 4 channelscorresponding to the different materials present: air, bone, healthytissue and tumor, along with 104 histories simulations of various dosecomponents (boron dose and dose due to neutrons and photons in differenttissues), as well as the error maps for each dose component. Given therelatively low number of histories, these simulations are susceptible tostatistical noise due to the algorithm´s incomplete convergence. Theneural network is trained to predict 108 histories simulations (whereconvergence is achieved) for these dose components, which are then used tocompute the final dose map. The L1 norm between the neural networkestimations and noise-free simulations obtained through the traditionalMonte Carlo approach serves as the cost function for parameter tuning.Figure 2 depicts examples of a patient´s CT scan, a statisticalnoise-corrupted Boron dose component simulated with low statistics, and anoise-free Boron dose component inferred by IA, respectively. Boron dosewas used as an example of one of the dose components inferred by theneural network for the sake of visual clarity.The proposed system underwent training using 80 different beam positionsacross 200 different patients from the Cancer Image Archive [2], resultingin a total of 16000 training instances. For testing, an additional 2000instances, corresponding to 25 patients, were employed. Importantly, noneof the patient data from the testing set was used during the trainingphase, ensuring the model´s generalization to unseen data. Simulationswere made using MCNP6.1 [3] and patients were modeled as 24x24x24 voxelarrays, where each voxel has sides of 1 cm.Across the testing set, 96.9% of the voxels in the 3D dose maps estimatedby the proposed system exhibited absolute differences of less than 5% ofthe maximum dose calculated in dose maps based on 108 histories MonteCarlo simulations. For reference, only 61.4% of the voxels of the 104histories dose maps fulfilled the same requirement. Furthermore, theproposed method achieved convergence with 104 less histories than theconventional approach, as inference time of the neural network isneglectable. This highlights its promise as a fast and reliable approachfor designing treatment plans within the context of BNCT.References:[1]: Kumada, H., & Takada, K. (2018). Treatment planning system andpatient positioning for boron neutron capture therapy. TherapeuticRadiology And Oncology, 2.[2]: Shusharina, N., & Bortfeld, T. (2021). Glioma Image Segmentation forRadiotherapy: RT targets, barriers to cancer spread, and organs at risk(GLIS-RT) [Data set]. The Cancer Imaging Archive.[3] J. T. Goorley. MCNP 6.1.1 Beta Release Notes. Los Alamos NationalLaboratory Tech. Rep. LA-UR-14-24680. Los Alamos, NM, USA. 2014.Graphic 1:https://icnct20.syskonf.pl/conf-data/icnct20/submissions/s31696504705_t4_v1_en.pngGraphic 2:https://icnct20.syskonf.pl/conf-data/icnct20/submissions/s31696504705_t5_v1_en.png