Researchers from Cardiff University and the University Hospital of Wales (UHW) have shown that combining their new system with existing AI tools can boost diagnostic accuracy by up to 1.5%, while also aligning machine decision-making more closely with expert radiologists’ judgment.

Their findings, published in IEEE Transactions on Neural Networks and Learning Systems, highlight how this approach could enhance diagnostic support for radiologists and encourage wider use of medical AI to help address challenges facing the NHS.

Dr. Richard White, Consultant Radiologist at UHW and clinical lead of the study, explained:
Computers are excellent at detecting abnormalities such as lung nodules by analyzing shape and texture. However, knowing where to focus on an image is a critical part of radiology training, with specific areas always requiring close review. This research combines those elements to explore whether computers can interpret chest X-rays in a way that more closely reflects how a trained radiologist works. This step addresses a gap in radiology AI research and is vital for building trust in AI-assisted diagnosis.

To achieve this, the team developed the largest and most robust visual saliency dataset for chest X-rays so far, capturing over 100,000 eye movements from 13 radiologists as they examined fewer than 200 images. This dataset was then used to train CXRSalNet, a new AI model designed to predict the regions of an X-ray most relevant to diagnosis.

Professor Hantao Liu, lead researcher from Cardiff University’s School of Computer Science and Informatics, added:
Current AI systems often lack transparency in how decisions are made – a major limitation in healthcare. Radiologists, on the other hand, draw on years of expertise and nuanced perceptual skills when examining images. By capturing where radiologists naturally focus their attention, our work uses eye-tracking data to ‘teach’ AI how to prioritize the most diagnostically important features in chest X-rays.

source: https://www.cardiff.ac.uk/news/view/2929144-knowing-where-to-look-researchers-create-ai-to-examine-medical-images-like-a-trained-radiologist