Deep Learning for Automated Cancer Care: An experienced clinical team can more than double a cancer patient’s odds of survival. In particular, radiation therapy (RT) is a key part of the treatment for 40% of all cancer patients with the potential to be highly curative and effective given the right treatment plan. However, a major barrier in that effectiveness is the expertise and experience of the patient’s clinical team in his or her particular cancer. We believe that AI can help bring the expertise of the world’s best clinical teams to every patient in every hospital, saving lives and costly resources. We have developed the world’s first AI-based treatment system that is being piloted to treat prostate cancer patients at Princess Margaret Cancer Centre today, with superior results to our standard of care for over 88% of patients. Cancer treatments are generated by the AI in minutes, instead of the hours-to-days taken under the typical manual process, increasing both efficiency and quality of care. Follow-up studies are starting for additional cancer types. Through a successful licensing deal of this Canadian technology to one of the world’s largest RT software companies we are also supporting deployment across Canada and around the world.
Machine learning for Intervention Quality Assurance: The QA process in RT involves ensuring that the RT plan is safe and effective. It is a manual process where a team of medical experts reviews the plan and associated segmentations (ROIs) and examines the amount of radiation to the target and organs at risk. Our work calculates image, ROI, and plan features, and uses a novel AI models to learn expected relationships between the features for clinically effective plans. A fusion between classification and density estimation enables novel plans to be classified as erroneous due to matching pre-existing training errors or by being an atypical plan, and thus worth clinical scrutiny.
Transfer and Meta Learning: Not all diseases are amenable to traditional machine learning with only hundreds of patients due to disease rarity or the rapid evolution of their related interventions. We are investigating approaches to learn features and structures from tens of thousands of patient images to enable application to rarer diseases with smaller datasets.