AI can boost radiation oncology clinical trials and clinical practice by providing better risk stratification, workflow and follow-up, according to experts who spoke at a panel discussion during Wednesday's plenary session.
(Left to right) Gensheimer, Feng, Li and Le
Discussion moderator Quynh-Thu Le, MD, from Stanford Medicine in Stanford, CA, said that AI is already in use in radiation oncology in the form of automatic segmentation, radiation treatment plans and quality assurance tools.
"Application of AI in radiation oncology is here to stay," said Dr. Le, chair of the Department of Radiation Oncology and the Katharine Dexter McCormick and Stanley McCormick Memorial Professor at Stanford University.
AI also has potential in clinical trials, an area where a large proportion of patients do not meet study criteria and many studies fail to complete recruitment on time. Dr. Le said that AI can help simplify eligibility criteria to expand the pool of eligible patients and provide more patients with access to trials. However, the need for continual updates will make it difficult to test AI tools in prospective clinical trials using the traditional designs.
"We need to come up with some novel designs and we need to work together with a new organizational framework to leverage the full power of AI," Dr. Le said.
Electronic medical records (EMR) provide a potentially rich vein of data that machine learning (ML) can mine for improved clinical decision-making, according to Michael Gensheimer, MD, clinical associate professor in radiation oncology at Stanford Medicine.
Recent research highlights the possibilities. Dr. Gensheimer shared results of a study that showed that an EMR-based algorithm did better than Stanford doctors in predicting short-term survival for patients with metastatic cancer.
Limitations Remain Barrier to Full AI Implementation
Data quality continues to be a key limitation. EMR records often fail to capture the date of a patient's death, meaning that studies using them could overestimate survival time. Documentation errors from patient-physician encounters are also plagued with errors—a problem that could be ameliorated with natural language processing that automatically extracts structured information from clinical notes.
Datasets also lack the necessary diversity to serve all patients. "We need to do better collecting data from underserved patient populations," Dr. Le added.
And then there is the issue of dataset sizes, something that can only be addressed through collaboration.
"There is a ton of medical data out there, but generally single institutions don't have that much," said Dr. Gensheimer. "We need to work on collaborating so that we can start to make our models better understand our patients' trajectories."
As a pathologist, Felix Y. Feng, MD, professor at the University of California in San Francisco, discussed using pathology-based deep learning (DL) tools to personalize treatment decisions, with a focus on prostate cancer.
Prognostic indicators are important to prostate cancer patients because treatment is based on the aggressiveness of the disease. A multi-modal AI tool developed using clinical data from more than 5,600 patients and more than 16,000 digital pathology slides has shown promise in this area, Dr. Feng said.
"I think the main take-home message here is that this AI prognostic tool outperforms standard risk stratifications approaches," he said.
The AI model also proved to be effective at predicting who will respond to androgen suppression therapy, a treatment for prostate cancer that carries side effects like impotence and osteoporosis.
Radiologic features of the tumor provide a rich source of information for AI-derived algorithms, said Ruijiang Li, PhD, associate professor of radiation oncology at Stanford Medicine.
Along with the commonly used metric of tumor size, AI algorithms use morphologic complexity and spatial variations between different regions of the tumor to risk-stratify patients.
A tumor's microenvironment plays a critical role in cancer progression and treatment response. Dr. Li suggested that a deep-learning algorithm could be trained to predict the status of this microenvironment through a CT signature of stroma cells, for instance.
"I think the future will be one in which we combine these different data sources and disciplines together in a multimodal framework that can really maximize our precision for predicting outcomes and guiding treatment decisions," Dr. Li said.
Access the plenary lecture, "Machine Learning in Radiation Oncology Clinical Trials and Clinical Practice," (W6-PL06) on demand at Meeting.RSNA.org.
