The use of Machine Learning System in Brachytherapy Treatment Planning

High dose rate (HDR) brachytherapy (BT) involves temporarily insertion of a sealed highly radioactive source inside or in close proximity of the target volume, by using catheters and/or applicators. Treatment planning in HDR BT is performed on a 3D set of computerized tomography or magnetic resonance images. The purpose of a treatment plan is to determine where (dwell position) and how long (dwell time) the radiation source should pause to expose its radiation. Each implanted catheter provides a set of possible dwell positions. Combinations of the possible dwell positions and dwell times, as well as utilizing contours of the target volume and organs at risk (OAR), contributes to optimization of treatment plans. The optimal plan has the best possible dose distribution with respect to the target volume and the OAR. Two key questions for treatment planning are how to evaluate a certain plan and how to search for the optimal plan. The traditional solutions to the two questions often lack the capability of fully utilizing previous experience, which result in inefficiency due to long computation times when solving the problems. In this thesis machine learning (ML) algorithms were investigated for evaluation and search for an optimal BT treatment plan. ML-based algorithms are chosen due to their ability to specialize in exploiting previous experience for better future performance. Evaluation of a certain treatment plan in the first problem requires calculation of the radiation dose. Clinical standards for BT dose calculation have traditionally been based on AAPM TG-43 report. In the TG-43 based dose calculation process, the affected malignant tissue, the surrounding radiation sensitive healthy organs, BT seeds, needles and applicators are considered to be water with unit mass density for simplification. This simplification overlooks the alteration of photon fluence and absorption of dose by different tissues, BT seeds, needles or applicators. Model based dose calculation algorithms (MBDCAs) provide a detailed and more accurate method for calculation of absorbed dose in heterogeneous systems such as the human body, with the Monte Carlo (MC) method being the gold standard. Recently, these algorithms have evolved from serving as a research tool into clinical practice in BT. To obtain accurate dose distributions, a correct geometrical description, density and tissue composition of the patient, a model of the BT seed and the implanted applicators with appropriate density and material composition are needed as inputs to these MBDCAs. AAPM TG-186 provides guidance for the use of MBDCAs. Although MC method is the most accurate technique for dose calculations, its use incurs an excessive computational cost and time. To provide a solution for the accuracy-time trade-off, a deep convolutional neural network (CNN) algorithm to predict dose distributions calculated with the MC method has been proposed in this thesis. The developed deep CNN based dose calculation algorithm was shown to be a promising method for accurate patient specific dosimetry in BT, at accuracy arbitrarily close to those of the source MC algorithm but with much faster computation times. Treatment plan optimization is traditionally solved as constrained optimization. Extended from core setup of plan optimization, a related plan analysis problem is formulated to focus on understanding the impacts of individual dwell position to overall plan performance. Derived from linear programming formulation of the optimization problem, reinforcement learning (RL) was applied to solve the plan analysis problem. Relying on the dose distribution at each dwell position, the RL solution showed flexibility in problem formulation as there is no need to enforce linearity. Moreover, when used in a simplified proof-of-concept setup derived from the real clinical case, the fully-trained RL agent showed the capability to reach optimal treatment plan within one step from any random start point

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.