A breakthrough machine learning model developed with the Vector Institute sets a new standard for cancer detection.
Researchers at the Temerty Faculty of Medicine, in collaboration with the Vector Institute for Artificial Intelligence, have developed a machine learning model capable of detecting early-stage cancers with 97 percent accuracy. The tool, described in a study published in The Lancet Digital Health, represents a major advancement in the application of artificial intelligence to cancer screening and has the potential to transform how cancers are detected in routine clinical practice.
The model, named OncoScan-AI, was trained on a dataset of over 2.5 million medical images and pathology reports from 14 Canadian hospitals. It is designed to analyze standard medical imaging, including mammograms, CT scans, and pathology slides, and flag potential early-stage malignancies that might be missed by conventional screening methods. In a validation study involving 50,000 previously screened cases, OncoScan-AI identified cancers that had been initially missed by human radiologists and pathologists in 23 percent of cases, while maintaining a false-positive rate of less than 3 percent.
"Early detection is the single most important factor in cancer survival, yet many early-stage cancers are missed during routine screening because the signs are incredibly subtle," said Dr. Elena Vasquez, the study's senior author and a professor of computational pathology at Temerty Medicine. "OncoScan-AI does not replace the physician; it serves as a second set of eyes that never gets tired and can process thousands of images with consistent accuracy. We see it as a tool that will make human experts even better at what they do."
The system uses a novel deep learning architecture that combines convolutional neural networks with transformer models, allowing it to analyze both the local features of individual cells and the broader patterns across an entire medical image. The model was designed with clinical workflow integration in mind: it produces a probability score and highlights the specific regions of concern within an image, providing radiologists and pathologists with actionable information that they can use to guide their clinical judgment.
The research team is now working with Ontario Health and several hospital networks to begin a prospective clinical trial that will evaluate OncoScan-AI's performance in real-world screening programs across the province. If the trial confirms the model's efficacy, the tool could be integrated into Ontario's breast, lung, and colorectal cancer screening programs within the next three to five years. The project has been supported by funding from the Canada Foundation for Innovation, the Ontario Institute for Cancer Research, and a $6-million philanthropic gift from an anonymous Temerty Medicine alumnus.
