AI systems that combine medical imaging, genetic data, and clinical records are helping oncologists predict which cancer patients will respond well to minimal treatment and which need more aggressive therapy, potentially improving outcomes and quality of life. This shift toward personalized cancer care represents a fundamental change in how artificial intelligence is being applied to medicine, moving beyond single-data-source analysis to integrated systems that understand the full complexity of individual patients.

Why Is Personalized Cancer Treatment So Hard to Get Right?

Every cancer patient faces a critical decision point: how much treatment is enough? Too little risks allowing the disease to progress; too much causes unnecessary side effects and reduces quality of life. For physicians, the challenge is prediction. Anne Martel, a professor at the University of Toronto and Vector Institute faculty member, focuses her research on exactly this problem.

Consider a patient with noninvasive breast cancer. Can AI predict whether they'll do well without radiotherapy, or whether they're likely to have more aggressive disease requiring additional intervention? These research questions directly affect patient outcomes and quality of life. Martel's work develops AI systems that extract actionable information from medical data to help answer these questions, focusing on medical imaging combined with clinical text data from patient reports.

How Are Researchers Building Better AI for Cancer Prediction?

The most exciting developments in medical AI involve what researchers call multimodal integration, combining multiple types of patient data into unified AI models. Martel's team recently presented work at the International Conference on Computer Vision (ICCV) demonstrating how to take large foundation models trained on digital pathology images and tune them using additional information from genomics, while also incorporating text from clinical records to guide what the model learns.

This multimodal approach allows AI to leverage complementary information that wouldn't be visible from any single data source alone:

• Pathology Images: Show tissue structure and cellular patterns that reveal how aggressive cancer cells appear under magnification.
• Genomic Data: Reveal molecular characteristics and genetic mutations that predict how the cancer will behave and respond to treatment.
• Clinical Text: Provide context about symptoms, treatment history, and outcomes that help the model understand the full patient picture.

The technical innovation involves adapting the model to handle multi-task learning, where classification tasks share similarities with survival prediction tasks. Rather than training separate models for each prediction goal, the system learns shared representations that apply across related tasks, effectively getting more insight from limited data.

What's Changing in How Medical AI Develops?

A decade ago, medical imaging researchers typically learned about new machine learning advances at general computer science conferences like ICCV, ICML, and NeurIPS, then adapted those methods to medical problems. The innovations flowed primarily from computer science into medicine. That dynamic has fundamentally shifted.

Today, novel developments increasingly originate within medical imaging research itself. The U-Net architecture, now ubiquitous for image segmentation across many domains, was originally developed for and presented at MICCAI, the main medical imaging conference. Medical imaging researchers are now creating foundational methods, not just applying them. The field has matured from following methodological advances to contributing new approaches that influence the broader AI community.

"I'm very much on the applied side. So you kind of have to ask yourself, well, this is great from a theory point of view, but is it ever going to be any use and is it appropriate? And also, have they tested it properly in a clinical domain?" explained Anne Martel, Professor at the University of Toronto.

Anne Martel, Professor, University of Toronto and Vector Institute Faculty Member

This shift matters for institutions supporting domain-specific AI research. Supporting medical imaging AI isn't just about applying general techniques to new problems; it's about recognizing that deep engagement with domain challenges often produces genuinely novel methodological innovations that benefit the entire field.

How Can Researchers Balance Innovation With Real-World Clinical Needs?

Martel's work sits at a productive tension familiar to many researchers working in applied domains: the pull between developing novel AI methods and delivering solutions that work for real problems today. Her approach balances both. For established clinical questions, her team applies and rigorously tests existing state-of-the-art methods, establishing benchmarks and demonstrating what's possible with current approaches. Simultaneously, they develop new methodologies designed to push beyond current limitations and improve accuracy.

This practical orientation reflects lessons from her experience co-founding Pathcore, a digital pathology company, in 2006. That experience provided valuable perspective on the difference between academic innovation and commercial viability. Research interests don't always align with market needs, and what excites researchers technically isn't always what drives adoption. For academics considering translation of their work, Martel suggests that understanding your own strengths and finding complementary partners matters more than trying to fill all roles yourself.

The progression from academic research to clinical application requires infrastructure that bridges computer science and medicine. Institutions like the Vector Institute facilitate this bridging by bringing together researchers from computer science and various application domains, creating natural opportunities for collaboration and knowledge exchange. For medical imaging researchers, such institutions provide access to AI engineering expertise that can help implement and scale methods, along with computational resources that make training large models feasible.

What Does the Future of Medical AI Look Like?

The research direction Martel and her colleagues are pursuing represents where medical AI is heading: away from isolated models that analyze single data types, toward integrated systems that combine imaging, genomic, and clinical information to build comprehensive understanding of individual patients' conditions. This shift promises more accurate predictions, better-informed treatment decisions, and ultimately, improved patient outcomes across oncology and beyond.