COMPASS

Medical AI research is thriving: over the last 30 years, more than 20,000 papers have been published in the field. This raises an important question: what problems is the research community actually addressing, and how are these problems chosen? Are researchers focusing on the areas of greatest clinical need, or are they selecting topics based on data availability, computational feasibility, or potential for technical novelty?

There is growing evidence that data availability, rather than clinical need, shapes research priorities. For example, 59% of 188 studies in a recent analysis of surgical AI focused on cholecystectomy [Carstens et al., 2025], a high-volume procedure that, however, is known for low complication rates. Similarly, the release of prominent lung imaging datasets and challenges was followed by a disproportionate increase in AI research on lung imaging [Varoquaux & Cheplygina, 2022]. In other words: data drives research.

Rather than letting data availability drive research (bottom-up), we want to invite the global clinical community to define the oncological questions where medical imaging AI could have the greatest transformative potential for clinical practice, i.e., problems where better AI capabilities could plausibly improve clinical decisions, workflows, or patient outcomes.

We have initiated a clinically grounded top-down approach to identify the oncological questions where medical imaging AI could have the greatest impact. Instead of perpetuating tasks defined by convenience or historical precedent, COMPASS places the clinical relevance of medical imaging tasks front and center. We aim to assemble a collection of clinical questions in medical imaging that, if solved with the help of AI, would meaningfully improve clinical practice for patients and caregivers. By making these priorities explicit, we aim to help steer research efforts and resources toward high-impact real-world problems rather than tasks primarily driven by convenience or academic interest

The identified challenges will cover different parts of the oncological workflow, from cancer prevention and screening to diagnosis, therapeutic decision-making, and follow-up care. They will be prioritized based on clear criteria of relevance, including their potential to improve patient survival, enhance quality of life, reduce healthcare costs, and mitigate clinical workforce shortages. To identify and define these clinical priorities, we will conduct a rigorous, global expert consensus process, bringing together perspectives from thought leading clinicians, AI researchers, ethics experts, regulatory stakeholders, and patient representatives.

Our ultimate goal is to reorient the efforts of the global medical imaging AI research community toward clinical utility and catalyze the next generation of meaningful AI advances in medicine.