Project Summary/Abstract Brain metastases (BMs) are a life-threatening disease, occurring in up to 40% of cancer patients. About 40% of BM patients have multiple (≥4) BMs (mBMs). Whole brain radiotherapy (WBRT), which has long been the standard of care for mBMs patients, has shown pronounced impairment of neurocognitive functions. Stereotactic radiosurgery (SRS) has improved tumor control and reduced negative effect on cognition function, compared to WBRT. However, it has been historically reserved only for patients with <4 BMs. Recently, several clinical trials reported strong evidence to support SRS for mBMs patients. National Comprehensive Cancer Network guidelines hence no longer restrict the number of BMs for SRS. However, the larger BM number in mBMs patients substantially increases the complexity of treatment planning. Conventional manual forward planning to manually determine plan parameters becomes cumbersome and impractical for mBMs. Modern inverse planning methods can determine plan parameters by solving an optimization problem that is composed of multiple objectives designed for various clinical or practical considerations, while the priorities among these objectives affect the resulting plan quality. The physician’s preferences for a particular patient can hardly be quantified and precisely conveyed to the planner, especially for mBMs patients due to the varying number, size, and locations of BMs. Hence, the best physician-preferred plan is often achieved through extensive trial-and-error priority tuning and several rounds of interactions between the planner and physician. Consequently, planning time can take up to hours, and plan quality may be suboptimal and can vary significantly, due to the varying levels of physician and planner’s skills and physician-planner communication and cooperation, leading to deteriorated clinical outcome. Inspired by the recent colossal advancements of artificial intelligence (AI), particularly deep reinforcement learning (DRL) and deep inverse reinforcement learning (DIRL), on intelligent decision-making in computer visions and robotics, we propose to develop an artificial intelligence driven automatic SRS treatment planning system for effective management of mBMs (Aid-mBMs), learning a human-level intelligence on treatment planning from human experts. We envision the system to have two deep neural networks: DNN-R that acts as an AI-physician to predict the physician’s preferences for each individual patient, and DNN-P that acts as an AI-planner to tune the priorities to achieve a plan of physician’s satisfaction. We will pursue two specific aims. Aim 1. System prototype development: We will collect human expert planners’ priority-tuning actions and develop DNN-R and DNN-P via interleaved DIRL-based reward function learning and DRL-based policy learning. Aim 2. System improvement and end-to-end evaluation: We will perform a prospective study to improve our system based on human expert’s further fine-tu...