Βy Nickolas Papanikolaou [Principal Investigator in Oncologic Imaging at the Fundação Champalimaud]
Researchers are using artificial intelligence to improve how prostate cancer is detected and diagnosed through MRI scans. In one part of the project, the goal was to identify whether a cancerous lesion is present in an MRI. Scientists from FORTH explored different AI strategies, including models that learn without labels, traditional deep learning, and approaches that segment the actual lesion area. Meanwhile, researchers at the Champalimaud Foundation experimented with cutting-edge techniques like self-supervised learning, where models learn patterns in the data without needing as much human input. Both teams also looked closely at how the size of the training data affects performance.
The second, and larger, part of the project focused on what’s called a “virtual biopsy” — using AI to predict the severity of prostate cancer without the need for invasive procedures. The team at Champalimaud developed models tailored to different MRI scanner types, combining deep learning with detailed image analysis (radiomics). They also investigated which features — from scan details to clinical and biological markers — best support accurate predictions. Importantly, they tested how well the AI could understand when it might be unsure about a result and whether this uncertainty could be used to avoid wrong predictions.
Other partners joined in as well. CNR explored how the area of the MRI image used by the model affects its accuracy, while FORTH compared how manual versus automated organ segmentation influences results. Across all of these efforts, researchers examined how much data is really needed to train reliable AI systems, helping pave the way for smarter, safer tools in prostate cancer care.
Champalimaud also trained models for biopsy reduction – in essence, these require a well performing solution capable of determining whether an individual might not have clinically significant prostate cancer without missing those that do. Initial studies showed this approach could reduce the number of unnecessary biopsies by 20% when compared with standard of care.
For other endpoints, model development was tricky: while relatively abundant when compared with other studies, data was scarce. Machine-learning models require large cohorts of data – typically thousands of cases – and collecting sufficient data for them was challenging. In the end, while tentative models were developed for other use cases, these did not lead to well performing solutions.
Finally, there was a continued effort at Champalimaud Foundation to biologically validate lesion detections. These detections come from AI models capable of telling where in an image prostate cancer is. Using prostates from patients undergoing surgical removal treatment, they compared how lesions appeared in the tissue with how they appeared in the magnetic resonance.
