Text sensitivity varies wildly across different medical image datasets.

Doctors use artificial intelligence models that combine medical scans with clinical text to outline diseased areas. But researchers do not know how much the text actually helps the computer pinpoint those pixels. To find out, the authors built a specialized decoding tool. It separates the evidence gathered purely from the image from the evidence influenced by the written notes. They discovered that different fusion methods barely change performance. More importantly, the model relies on text very differently depending on the dataset. On two medical datasets, removing the text caused huge performance crashes. On two other datasets, the text barely mattered at all. Finally, the researchers found that text works by shaping overall meaning across the whole picture, rather than pointing out exact physical locations.

Why it matters

This matters for medical AI developers and clinicians because understanding when text descriptions are crucial ensures models use clinical knowledge appropriately. If text is unnecessary for some tasks, models can be simplified; if essential, relying solely on images could lead to failures.

Questions answered