Monday, November 30 | 9:30 a.m.-10:30 a.m. | M3-SSIN02-2 | Room E450B
In this scientific session on imaging informatics, a study suggests that expert radiologists who scan more broadly when interpreting chest x-rays achieve meaningfully higher diagnostic accuracy than those with narrower patterns.
In an analysis of 103,592 chain-of-thought reasoning entries from 501 radiologists across 71 countries interpreting 50,312 chest radiographs, researchers examined clinical context integration, visual search trajectories, inter-reader agreement, and uncertainty expression patterns using the CheXthought dataset, a global collection of synchronized radiologist reasoning and visual attention data.
Visual attention trajectories clustered into three patterns, according to the findings: broad (35.5% of interpretations, covering 11.7 anatomic zones, 78.2% accuracy), central (32.3%, 10.0 zones, 73.6% accuracy), and narrow (32.2%, 6.4 zones, 68.3% accuracy). Clinical context was referenced in 54.9% of interpretations and was associated with a 4.5-percentage-point accuracy gain overall, with the largest benefit for pneumothorax (+16.0 percentage points), cardiomegaly (+5.9 points), and lung opacity (+3.9 points). Inter-reader agreement was highest for pneumothorax (kappa = 0.92), pleural effusion (0.90), and cardiomegaly (0.89), and lowest for pneumonia (0.30) and pleural other (0.23). Uncertainty was expressed in 84.1% of interpretations for suboptimal-quality studies versus 44.2% for optimal studies, and was most frequent for pneumonia (78.1%) and consolidation (75.1%).
Ultimately, a granular understanding of how expert radiologists integrate clinical context, scan images, and express uncertainty is considered essential for building AI systems that replicate expert cognition, the authors noted.
“These reasoning patterns provide a framework for training AI systems that communicate uncertainty, adapt to image quality, and replicate expert visual search strategies in clinical practice,” according to the group.
Take in this Monday morning presentation to learn more.



