Analytical Models and Data-Driven Methods for Radiation Therapy Treatment Planning
The clinical goal of radiation therapy (RT) is to maximize the tumor damage and kill all the cancerous cells while minimizing toxic effects on surrounding healthy tissues during the course of treatment. Adaptive radiation therapy (ART) has been widely used to adjust the radiation dose in response to potential changes in tumor volume during the treatment to reduce the radiation toxicity in healthy organs. One of the key challenges in ART is to determine the best time to adapt the plan in response to uncertain tumor biological responses to radiation during the treatment. Tumor biological response change dynamically over time and can be different from one patient to another. Therefore, considering tumor biological responses to radiation in ART treatment planning is challenging due to the high levels of uncertainty in biological factors. Determining the possibility of treatment side-effects for each patient before starting the treatment is another challenge in radiation therapy treatment planning. This dissertation focuses on a combination of optimization, deep learning, and statistical methods to address the aforementioned challenges in this field and improve the survival of cancer patients treated with radiation therapy. We will tackle this problem from two different perspectives: (1) developing effective personalized radiation therapy treatment plans and (2) predicting possible critical side-effects of the treatment for each patient before the treatment. First, we propose an automated radiation therapy treatment planning framework using reinforcement learning (RL) which incorporates uncertainty in tumor biological responses during treatment to find the optimal policy for ART. We also provide a novel tumor response model to estimate tumor volume changes and radiation responses during the treatment. This approach helps the decision-maker to control both biological and physical aspects of the treatment and achieve a robust solution under biological uncertainties without dealing with complex optimization models. The presented method provides much-needed flexibility in which a plan can be customized based on the patient case, cancer type, and the decision maker’s preference on treatment outcomes. Second, we address one of the critical radiation therapy treatment side-effects known as radiation-induced lymphopenia (RIL). RIL occurs due to a severe reduction in the ab solute lymphocyte count (ALC) after radiation exposure and can seriously affect patient survival. Therefore, we aim to assess the role of radiation therapy in ALC depletion to determine high-risk patients. To accomplish this goal, two mathematical models are proposed to approximate lymphocyte depletion based on radiation dose distributions and the ALC baseline for radiation therapy patients. Finally, we compare the potential post-treatment lymphocyte survival outcomes in cancer patients for photon and proton-based RT modalities. Third, we develop a hybrid deep learning model in a stacked structure to predict the ALC depletion trend throughout radiation therapy treatment for cancer patients based on the pretreatment clinical information. Then, we extend the model to account for making predictions after the initial phase of treatment (e.g., at the end of week 1). A discriminative kernel is also developed to extract and evaluate the importance of temporal features. The presented deep learning structure can efficiently use information from different groups of clinical features to predict ALC depletion without requiring a large amount of data to process too many features while reducing bias and generalization error. This approach helps the physicians to identify patients at risk of severe RIL who might benefit from modified treatment approaches which ultimately improve survival of the patients. In the last part of this dissertation, we provide an approach to estimate prediction intervals for ALC values. The proposed approach enables practical implications of predictive models in clinical decision-making by estimating the individualized predictive uncertain ties. Finally, a comprehensive hybrid decision-making framework is proposed to assess RIL risk for a given patient based on a given treatment plan and its predicted post-treatment lymphocyte survival outcome. This decision-making framework can be used as a guide for physicians to take advantage of advanced deep learning models and make appropriate decisions in selecting the safest treatment plan for an individual patient in clinics.