It explains AI medical diagnoses by altering images using only the classifier's internal logic.

Doctors hesitate to use deep learning for medical diagnosis because they cannot see how the computer makes decisions. One way to test a model is with counterfactual imagescounterfactual imagesModified images that show what would cause the AI to change its diagnosis.. These show how a scan must change to flip the model's diagnosis. Usually, researchers make these using complex second tools like generative adversarial networksgenerative adversarial networksAI models that create new images by having two networks compete. or diffusion modelsdiffusion modelsAI models that generate images by gradually refining noise.. But using a second black-box model muddies what the first model is actually seeing. To solve this, researchers built a method that needs no generative model or extra training. Instead, it extracts causal evidencecausal evidenceDirect information from the AI that influences its predictions. straight from the classifier to make deterministic edits in specific regions of interest. On real-world medical datasets, this altered predictions while staying closer to the original images than generative tools.

Why it matters

Clinicians need to trust AI diagnoses, but current tools are hard to explain. This method makes the AI's reasoning more transparent, helping doctors and patients feel confident in the technology.

Questions answered