Bayesian Optimization in Multi-Objective Treatment Plan Optimization of High Dose Rate Brachytherapy

Along surgery, chemotherapy and immunotherapy, radiation therapy is prescribed to more than 50% of the cancer patients.High dose rate (HDR) brachytherapy is a radiation-therapy modality that is often prescribed to treat genitourinary, lower gastrointestinal, and breast cancers.In this modality, the radiation source is placed near or inside the tumors via hollow catheters.Treatment planning in HDR brachytherapy is the process of optimizing the radiation dose to the tumor and organs at risk by finding the time and the location where the radiation source should dwell.In order to find the optimal configuration of dwell times for a given tumour site, an inverse planning optimization algorithm is required.To optimize the dwell times, a medical physicist manually fine tunes a set of hyper-parameters known as the penalty weights.Selection of organ penalty weights to obtain clinically acceptable dose volume histogram (DVH) metrics in inverse treatment planning process of HDR brachytherapy is a time-consuming task and plan quality depends on planner's experience.The penalty weights represent the relative importance of absorbed dose planned to be delivered to various structures in the objective function of the dwell time optimizer.Therefore, the penalty weights directly control the trade-off between DVH metrics of structures.Obtaining these trade-offs (Pareto optimal solutions) is an active area of research as automation of this process is essential to fully automate the treatment Abstract ii plan optimization process.In this thesis, we implemented two algorithms that perform penalty weight tuning retrospectively on data for 10 prostate cancer patients.The first one is a simple interval shrinkage heuristic called semi continuous interval shrinkage exploration (SCISE), which evaluated hundreds of penalty weights in parallel and established the effect of the range of the penalty weights on the DVH metrics for each patient.The second algorithms was multi objective Bayesian optimization (MOBO) with q-noisy expected hypervolume improvement (qNEHVI) as its acquisition function.MOBO-qNEHVI aims to find Pareto optimal solutions to any multi-objective black box function with as little function evaluations as possible.In this study, fast mixed integer optimization (FMIO) was used as dwell time optimizer to evaluate the fulfillment of DVH objectives for each penalty weight vector.SCISE showed that the relative magnitude of the penalty weights with respect to one another controls the plan quality, not the specific magnitude of the weights.Therefore, we fixed the weight of the PTV to 1 and defined the range for OAR weights to be from 0.001 to 1.Then, we used MOBO-qNEHVI on this configuration and tested the impact of the number of MOBO-qNEHVI iterations and the range of penalty weights for OARs on the fraction of treatment plans that are clinically acceptable (acceptance rate) and as well as on the performance time.The acceptance rate of MOBO-qNEHVI increased logarithmically with increasing number of FMIO calls, while the performance time increases exponentially.Initializing the surrogate Gaussian function with random weights improved the acceptance rate of MOBO-qNEHVI for low number of iterations.MOBO was able to find patient specific Pareto optimal weight combinations for all patients with average success rate of 80.0 ± 13.9% and performance time of 1.65±0.39minutes.Since no training data is required, MOBO-qNEHVI Abstract iii can be applied to other cancer types that are treated with HDR brachytherapy.Therefore, this method is a reliable and general multi-criteria optimizer in HDR brachytherapy.Abrégé vi avec l'augmentation du nombre d'appels FMIO, tandis que le temps de performance augmente de façon exponentielle.L'initialisation de la fonction gaussienne de substitution avec des poids aléatoires a amélioré le taux d'acceptation de MOBO-qNEHVI pour un faible nombre d'itérations.MOBO a pu trouver des combinaisons de poids optimales de Pareto spécifiques à chaque patient avec un taux de réussite moyen de 80,0 ±13, 9% et un temps de performance de 1,65 ±0, 39 minutes.Puisqu'aucune donnée de formation n'est requise, MOBO-qNEHVI peut être appliqué à d'autres types de cancer traités par curiethérapie HDR.Par conséquent, cette méthode est un optimiseur multicritère fiable et général en curiethérapie HDR.This work was supported by Canada Research Chair (grant number 252135) as well as the Canadian institute of health research (grant number 103548).Computations were performed on the Calcul Québec Beluga and Mp2 clusters.The operation of these supercomputers is funded by the Canada

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.