AI models for medical imaging usually learn to match a scan directly to a disease name, so they fail on diseases never seen in training and cannot explain their decisions in a doctor's terms. Radiologists instead first identify visible findings, like how dense a lesion is or how it absorbs contrast, and then combine those findings according to published guidelines to name the disease.
The researchers built a model that learns to recognize individual findings by matching image regions to text descriptions, and then composes a disease from the findings its guidelines require, without learning the disease itself. This makes the model interpretable and lets it handle new diseases by simply listing their findings.
On multi-phase liver computed tomography (CT) scans, this approach spots and localizes diseases never seen in training significantly better than all comparators, including rare diseases and cases from another institution. It also improves diagnosis in the fully supervised setting, showing that findings-first learning generalizes better than disease-first learning.