A research team from LMU Munich, the LMU Hospital, the Karlsruhe Institute of Technology, and the University of Bayreuth investigated how different forms of AI support influence diagnostic work in radiology. The result: Particularly explainable, step-by-step AI explanations significantly improve diagnostic accuracy.

In the study, 101 radiologists analyzed real clinical cases based on CT and MRI images. The participants either worked without AI support or received hints from a multimodal language model – in the form of a simple diagnosis, a differential diagnosis, or a so-called "Chain-of-Thought" explanation.

Step-by-step explanations increase the diagnostic accuracy of doctors

It was found that doctors achieved the best results with step-by-step AI explanations. The diagnostic accuracy was significantly higher than the control group without AI support. Particularly helpful was the AI's ability to transparently explain its conclusions and clearly present image features, clinical hints, and exclusion criteria.

The researchers emphasize that not only the quality of the AI response is crucial, but also its explainability. Pure diagnostic suggestions or lists of possible diagnoses can instead lead to so-called automation bias – where users adopt AI recommendations without verification.

The results underline the importance of explainable AI systems in healthcare. AI is not intended to replace medical expertise but to serve as a supportive tool that makes decisions more transparent and better verifiable.

According to the researchers, the insights are also relevant beyond medicine: For the successful use of generative AI, not only the performance of the models is decisive, but also the quality of the interaction between human and machine.