Artificial Intelligence for Prostate Cancer Treatment Planning

PHASE 1

1. Develop a process and tools (DAST) to capture the rationale, criteria, and logical basis behind the treatment planning process using well understood Human Factors knowledge gathering methodologies and Machine Learning tools.
2. Build the AI technology to learn the process and apply it to generating treatment plans. Images and expert-drawn volumes from Radiation Therapy Oncology Group (RTOG) 0938 will be used for initial training of the AI system. These data are not Dartmouth-Hitchcock Medical Center (DHMC) patients, but rather were consented and acquired through RTOG 0938. These data are housed at NRG/RTOG headquarters in Philadelphia. NRG already has an established, IRB-approved protocol for exploring AI systems for the 0938 data set. A minimum of 30 cases will be used for this initial training work for the AI system to "learn" volume segmentation of important structures and targets. Additional patients from the available 200+ patients on 0938 may be added for additional AI learning of volume segmentation as initial software programming is implemented.
3. Determine the optimal number of historic treatment plans to train the AI technology and test it. 45 cases will be provided by NRG from NRG/RTOG studies 0415, 0126, and 0521 (totaling 135 cases), respectively representing favorable-, intermediate-, and high-risk prostate cancer treatments. These scans and expert-defined volumes are part of NRG datasets housed in Philadelphia. All these patients already signed study-specific consents which included permissions to allow personally specific clinical information to be used for other cancer-related studies. IRB review of the use of these data for this specific study is anticipated to be achieved through NRG mechanisms in the context of these prior consents, to assure that the use of these data for this specific AI study is appropriate and approved. Provision is anticipated of a total of 30 additional patient cases as "gold standards," 10 each from DHMC, University of Massachusetts (UMass), and Oregon Health Sciences University (OHSU), respectively, all initially planned and treated within the 2015-2018 time frame. The first 5 from each institution will be favorable-risk patients, and the next 5 from each institution will be high-risk patients (thereby achieving a wide range of treatment approaches, with more to be added subsequently). As part of this effort, each individual institution will contact its own specific patients (5 favorable-risk, 5 high-risk) to obtain study-specific consent for the use of their data for this protocol. Once anonymized, these scans and plans will be shared across all three institutions. For each of these patients, the other two ("non-host") institutions will create their own volumes and plans. Using a modified Delphi approach, the three teams will then meet to generate an agreed-upon "composite" plan for each patient. Thus, in total there will be four treatment plans for each of these 30 patients, yielding 120 total plans that will serve as the "gold standard" for this AI project, and will be inputted for testing/validation to the AI system. PHASE 2
4. Expand the database to include intermediate-risk patients, 5 respectively from each institution, following the above procedures, to yield an additional 60 plans to serve as additional inputs for the AI system.
5. Validate and test the AI technology by inputting patient images and target delineations from historic case data and assessing whether the AI technology-generated plans are "consistent" with the final plans that were created by expert clinicians.
6. Test the technology with new patient case data and validate the plan with a team of expert clinicians. This will involve "modified Turing tests," as developed in NRG-RTOG studies exploring AI applications.