New Method Cuts AI Hallucinations in Medical Imaging Reports

Large language models trained on images and text have made medical report writing faster and more fluent, but they suffer from a critical weakness: they often invent plausible-sounding diagnoses that have no basis in the actual scans. A new approach called Layer-wise Expert-aligned Decoding (LEAD) aims to solve this by working directly within the model's internal machinery to steer it toward factuality.

The Hallucination Problem#

Vision-language models excel at generating coherent, well-written radiology reports. However, their reliance on statistical patterns in training data can lead them to produce confident-sounding descriptions of pathologies that aren't actually visible in the input images. Prior solutions have attempted to counteract this by feeding external knowledge or guidance into the model, but these add fragility and don't address the fundamental mismatch between what the model "sees" and what it generates.

How LEAD Works#

The LEAD method takes a different approach. Rather than bolting guidance onto the model from outside, it intervenes within the generation process itself. The technique creates a specialized expert module designed to identify and extract clinically meaningful pathological features from medical images. These features are then fed into each layer of the model's decoder through a gated mechanism—essentially a learned valve that decides how much expert signal to inject at each step. This allows the large language model to progressively incorporate image-grounded information as it constructs each word of the report, helping it stay anchored to what the imaging actually shows.

Results and Implications#

Testing on multiple public datasets shows that LEAD improves both clinical accuracy and factual consistency without sacrificing the quality and readability of the generated reports. By addressing hallucinations at their source—in the model's decoding biases—the method offers a more robust path to trustworthy AI-assisted medical documentation than existing external-guidance approaches.