Optimizing Efficiency and Quality of Brachytherapy for Cervical Cancer using Machine Learning Based Automation

点击翻译按钮获取中文摘要

Current treatment planning for brachytherapy of cervical cancer is performed with manual techniques that are both time-consuming and subjective. Manual treatment planning takes 95 minutes on average and occurs while patients are sedated, and the quality of the treatments is highly dependent on the expertise of the physician. Unfortunately, the resource intensiveness and need for specialized expertise are barriers to implementation of brachytherapy, and as a result many centers are not offering this essential treatment for cervical cancer. Alarmingly, this rapid decline in brachytherapy utilization has been linked to 12% reductions in patient survival. To overcome the barriers to delivering highly effective brachytherapy, there is a critical need for tools that improve the efficiency and reduce the complexity of treatment planning for each patient. My long- term goal is to become an independent investigator focused on automating brachytherapy cancer treatment with machine learning, producing button-click solutions that will significantly upgrade the quality of brachytherapy and combat declining utilization. I have significant experience in modeling, image processing and computer programming and I want to build on this skillset with a training program that will prepare me for independence. I have assembled an exceptional mentorship team, which includes expertise in machine learning, clinical trials, implementation science and statistics. We formed a training plan to gain expertise in (1) deep learning, (2) advanced statistical analysis, (3) design of clinical trials and implementation of technology and (4) research career development. The research goal of this proposal is to develop a tool for fully automated cervical brachytherapy treatment planning, which uses machine learning models to make predictions for new patients. The central hypothesis is that automated planning using machine learning will generate non-inferior or even superior plans in significantly reduced treatment planning time. This hypothesis will be tested with the following specific aims: (1) Develop machine learning models, which use labelled patient images to predict radiation dose; (2) Develop and evaluate efficacy of a pipeline for automated brachytherapy planning; and (3) Prospectively measure the efficiency and clinical impact of automated brachytherapy planning. For Aim 1, convolutional neural networks will be developed to predict 3D radiation dose from imaging. Aim 2 will convert predicted doses into deliverable treatment plans using gradient-descent optimization to determine optimal treatment parameters. Aim 3 will provide an end-to-end validation of the automated planning by testing it in real-time clinical workflow. This work is innovative because it presents the first clinical validation of an automated treatment planning system for brachytherapy of cervical cancer. The proposed research is significant because it will revolutionize the current brachytherapy paradigm by applying machine learning to automate and standardize time-consuming, manual processes. This work is a key step towards my future R01 submission on multi-institutional implementation of automated cervical brachytherapy.