Improving initialization in MO-RV-GOMEA optimization for cervical cancer brachytherapy treatment planning

In High-Dose-Rate BrachyTherapy (HDR-BT), a form of internal radiotherapy, the time to construct a treatment plan is desired to be as short as possible. To reduce the trea tment planning time, a shift from manual to automatic treatment planning has been made over the past decades. Specifically, automatic treatment planning for HDR-BT is aimed to output treatment plans that satisfy a clinical protocol and are produced within a limited time frame. An optimization algorithm that is particularly well-suited for performing such a task is the Multi-Objective Real-Valued Gene pool Optimal Mixing Evolutionary Algorithm (MO-RV-GOMEA). Recent applications of MO-RV-GOMEA to prostate HDR-BT treatment planning have been shown to produce high-quality treatment plans efficiently. However, the application to cervical HDR-BT treatment planning has exemplified the need to further decrease the run-time required by MO-RV-GOMEA to produce a clinically viable treatment plan solution (evaluated by the so-called golden corner). We propose an ensembled Feedforward Neural Network model with Exponential search and Interpolation output correction (eFNN+ExpInt) to generate a set of treatment plan solutions that can be used to initialize MO-RV-GOMEA. The eFNN+ExpInt model’s main purpose is to resemble a proof-of-concept initialization that can reduce the required MO-RV-GOMEA optimization time toward clinically viable treatment plans. Our results show that the required MO-RV-GOMEA optimization time to reach a clinically viable plan is significantly reduced for the majority of tested patients by the use of eFNN+ExpInt initialization. Specifically, the required average runtime (over 25 patient fractions) to reach the golden corner was 10.662 seconds for eFNN initialization, 9.297 seconds for eFNN+ExpInt initialization and 11.156 seconds for Heuristic+ExpInt initialization as opposed to 21.597 seconds for the current initialization benchmark. A warm start initialization such as eFNN+ExpInt can enable MO-RV-GOMEA to converge faster to a clinically viable solution and can open opportunities for more sophisticated and time-costly patient treatment planning solutions.