Project Summary/Abstract The Bragg peak serves as the foundation for proton radiotherapy (RT), providing particular benefit to reduce treatment toxicity and improve quality of life for patients with cancers that are very close to multiple critical organs, particularly for pediatric patients. However, anatomical changes often occur during RT treatment course, caused by tumor shrinkage, patient weight loss, and daily treatment setup variations. Proton RT is very sensitive to anatomical changes, as a small density change along the beam path can shift the Bragg peak, resulting in significant underdosing of tumors and overdosing of adjacent critical organs. Adaptive proton RT (APT) offers a general solution to account for anatomical changes. However, current APT workflow is resource-intensive in terms of time and equipment and requires extensive work from the entire clinical team. It typically takes a few days from the CT image acquisition of patient’s new anatomy to the time when the new plan is ready for treatment, resulting in 1) interruptions of the treatment course, which can impair the local control and overall survival rate due to tumor cell repopulation, particularly for cancers of fast tumor repopulation; 2) time delays between the image acquisition of new anatomy and the treatment delivery of the new treatment plan, which can compromise the plan’s efficacy due to the continual anatomical changes occurring during the gap. Online APT is much needed to guarantee high treatment quality throughout the treatment course. While several commercial systems support online adaptation for photon RT, there is no such system available for proton RT. This gap exists because online APT requires superior contouring accuracy and more robust online adaptation algorithms that can account for large anatomical changes due to the Bragg peak's sharp dose fall-off and its acute sensitivity to anatomical changes. To fill this gap, we propose to develop an artificial intelligence-driven system for online adaptive and personalized proton therapy (AID-ON-APPT). We envision that this system will calculate the actual daily dose distribution of the treatment plan based on patient’s daily anatomy and treatment setup, quantify the actual plan quality before the delivery of each treatment fraction and, if needed, generate a new optimal plan within minutes. The proposed system will be developed via three specific aims, using head-and-neck cancer as a test bed: 1) Build two patient-specific deep-learning (DL) models for automatic online delineation of tumor and organs at risk (OAR) on daily cone-beam CT images; 2) Build a patient-specific virtual replanner network using deep reinforcement learning and prior knowledge for automatic online replanning; 3) Conduct a clinical study of virtual online APT on 100 HN patients to test system performance and its potential clinical benefits. Upon successful completion, the proposed platform will enable online adaptive and personalized proto...