- Development of Artificial Intelligence (AI) Tools to Understand and Duplicate
Experts’ Radiation Therapy Planning for Prostate Cancer
Fast-Track proposals will not be accepted.
Number of Anticipated Awards: 2-3
Budget (total costs, per award): Phase I: up to $300,000 for up to 9 months; Phase II: up to $2,000,000 for up to 2 years
PROPOSALS THAT EXCEED THE BUDGET OR PROJECT DURATION LISTED ABOVE MAY NOT BE FUNDED.
The goal of this topic is to stimulate Artificial Intelligence (AI) technology to improve treatment planning for prostate cancer by training algorithms to “read” standard Computerized Tomography (CT) and/or MRI images and recommend suitable treatment plan approaches. The resulting AI software may be a tool to aid radiation oncologists in reaching consensus treatment planning, reducing professional costs, and improving quality assurance in clinical trials and patient care. Also, by understanding the AI processes used to achieve an optimal solution, the software may have application in training junior radiation oncologists and updating practitioners.
Treatment planning for radiation therapy is becoming increasingly complex with the advent of image-guided radiation therapy (IGRT) and charged particle therapy (CPT). A substantial amount of physician time and effort is allocated to locating and contouring key tumor and normal tissue structures. Prostate cancer is chosen as it is a common disease worldwide, has well-defined risk groups and patient data-bases including outcomes. The treatment decision process involves assessing the patient’s risk status for disease progression based on tumor size, grade and biomarkers (e.g., prostate specific antigen or other tests); and assigning one of three risk groups: low, intermediate and high. Imaging includes CT and often MRI. Based on the images obtained, the physician and medical physicists plan the target volume to be treated and normal tissue to be avoided. In practice, treatment guidelines are established by consensus papers. However, even between world-renowned experts, treatment plans can exhibit significant differences.
It may be possible to go beyond verbal consensus texts as a basis for defining treatments using an approach similar to AI-based contextual image analysis that does not rely on an understanding of the rationale behind expert “preferences” in treatment plans. Such an AI-based approach would provide an agnostic initial plan, based on computerized image interpretation, upon which the physician and physicist could build a treatment plan. In addition, by studying the AI processes used to achieve an optimal solution, the processes for clinician decisions could be further optimized. The AI software delivered through this contract solicitation could reduce the time burden of image segmentation 75%, from four hours to one hour or less freeing up time for patient care. (https://www.technologyreview.com/s/602277/deepmind-will-use-ai-to-streamline-targeted-cancer-treatment). It may result in a substantial reduction in time for physicists and physicians and may improve quality control by having AI assist in initial plan. For smaller facilities with limited funds for staffing, this could improve quality by defining an initial plan developed by AI that could then be reviewed and modified by the physician without starting from unannotated images.
The goal of this contract solicitation is to develop and evaluate AI’s capacity to duplicate expert radiation therapy planning. The purpose is to develop radiation therapy treatment plans through AI interpretation of radiomic data from diagnostic images with the intent of fully or at least largely automating treatment planning to eliminate subjective biases, improve treatment quality and reduce cost. The objective of this FOA is not to achieve a breakthrough in Artificial Intelligence, but rather benefit from the recent advances in the development of treatment planning systems and machine learning to improve radiation therapy by eliminating repetitive, time-consuming and subjective biases in treatment delivery, which can result in sub-optimal plans and inadvertent normal tissue injury.
The initial goal is to improve the outcome for patients with prostate cancer. By developing knowledge-based planning solutions it may be possible to provide a more standardized treatment, which would facilitate quality assurance, possibly extending it to facilities with limited expert personnel, and facilitate the conduct of research by reducing the variability and apparent arbitrariness and/or preference that individuals incorporate in their treatment design. The long-term goal will be to apply this technology to other tumor sites.
There are many considerations that go into the selection of a target volume for treatment. Nowadays prostate cancer radiation therapy is based on “risk” stratification groups (low, medium and high), which generally determine the tumor dose, volume and other ancillary treatments such as hormonal therapy. Thus, the target volumes include the prostate, varying amounts of the seminal vesicles and the local lymph nodes for the more advanced risk group. The normal tissues are the rectum, particularly the anterior rectal wall, base of the bladder, femoral heads and occasionally additional abdominal content for lymph node fields.
There are emerging algorithms being developed to outline the normal tissues and the prostate. The scope of the activities here would be having three world renowned experts outline the same set of cases of the varying risk group with the process being “watched” by AI. The expert would dictate the thinking of why the chosen treatment volume and dose are being selected and this would be transcribed. Enough training cases would be used (the estimate of training cases needed is part of the proposal) for the AI to then take a second “test” batch of patients being planned and compared how the AI does in comparison to each of the experts. One question would be how many training cases it takes for the AI to reliably anticipate what the expert will do and to understand the discrepancies between the expert and AI.
Next, using the results from the second “test” batch, the plans for the three experts will be compared, as in consensus panels, and the plans by the AI system for each of the experts will be compared to see if the AI “understood” the differences and how AI would reach a consensus. Should this be effective, one could begin to use the AI to do the initial plan. Some of the cases could be chosen that had grade 2 bowel or bladder side effects to see how the AI and expert plans approach this.
Projects That May Be Supported:
Algorithms for AI are now rapidly emerging. This proposal would allow small businesses and start-ups, often comprising the most creative new people in a field, to test their creativity solving a clinical problem that has some degree routine and repetition. The support would be used for assembling and anonymizing the treatment planning, supporting some of the time and facilities of the experts and for bringing them together for consensus discussions. The AI group would receive support for time and resources. AI platforms such as Google TensorFlow, IBM Watson or Definiens’ Image Intelligence suites are likely to be used to emulate human cognitive process of treatment planning and then extract information and develop AI algorithms.
Projects That Will NOT Be Supported:
Proposals from the large manufacturers. AI that only outlines tumor and normal tissues but does not select a treatment plan for the three risk groups.
Phase I Activities and Deliverables:
Design and deliver an AI approach to develop radiation therapy planning for prostate cancer.
- Choose three expert radiation therapy planning teams comprised of a physician and planners (i.e., a person who is knowledgeable in treatment planning with good understanding of the treatment planning system) and evaluate expert cognition process in developing treatment planning for all three strata of patient risk groups (i.e., low, intermediate, and high).
- Teams including experts must be identified prior to submission of the proposal.
- Companies must already have a dataset including patients in the 3 risk groups, including radiation data and outcome (at least one year post treatment to assess toxicity) in hand before Phase I starts. Additional resource datasets that they can use to test the performance of their auto-segmentation tools can be the annotated prostate database from TCIA (The Cancer Imaging Archive) : https://wiki.cancerimagingarchive.net/display/DOI/NCI-ISBI+2013+Challenge%3A+Automated+Segmentation+of+Prostate+Structures as well as the non-annotated dataset: https://wiki.cancerimagingarchive.net/display/Public/SPIE-AAPM-NCI+PROSTATEx+Challenges.
- Identify criteria by which an expert planner develops each treatment plan and plan for each risk group
- Describe plan on harmonizing imaging for tumor and normal tissue identification
- Present a justification for the number of training and validation sets that would be needed for each of the risk groups so that the AI results can provide a starting point for the planning team to refine the initial plan and determine the final course of treatment. (If this had been underestimated, it is expected that a suitable number of additional cases will be obtained).
- Design and develop computational methods aimed at developing treatment planning for prostate cancer patients.
- Propose plan to develop, incorporate, compare AI planning system with expert treatment planning system and validate AI based treatment planning system
- Present AI concept to develop knowledge-based radiotherapy treatment planning to SBIR Development Center and the Radiation Research Program
- At a minimum, this technology should be applied to standard 3D CT datasets. Use of additional imaging is at the preference of the planning team, which could include MR fusion into the planning CT to define size/shape of the gland rather than using CT alone.
- Describe and demonstrate cross validation of image delineation, reproducibility of the Planning Target Volume (PTV) and treatment plans.
Activities and deliverables that will be used to evaluate whether the project should continue to be funded for Phase II include:
- Creation of an algorithm.
- Demonstration of the ability for the AI to provide a treatment plan (or 3 options of plans).
- Estimation of the number of cases needed to compare the verbal consensus by the three planning teams and the consensus by the AI from each of the teams.
- Concordance between expert treatment plans and AI plans.
- Use of datasets for training, testing, and validation.
- Execution and validation of computational tool, method, or model.
- Establishment of partnerships for potential empirical validation.
Phase II Activities and Deliverables:
- Refinement of algorithm.
- Demonstration of utility of AI plan as the initial step to then be reviewed and modified by the planning team.
- Apply AI to data sets and determine how many sets are required before physician and AI are largely in agreement.
- Expand types of data sets to include MRI or PET or other sources of information that would improve AI’s performance.
- Establish external partnership(s) for empirical validation of method, as demonstrated with letters of intent from strategic partners.