Radiologist gaze panels improve chest X-ray severity prediction accuracy
GazeRIB: Systematic Evaluation of Multi-Reader Gaze-Guided Chest X-ray Learning
Computer Vision and Pattern RecognitionHuman-Computer Interaction
Summary
Looking where experts focus on chest X-rays can help teach computers to better understand the images. The authors studied how using groups of radiologists’ gaze patterns instead of just one expert can improve predictions of disease severity. They found that combining gaze information from three or more experts leads to better computer models than using one alone. However, the mix of experts and the specific patient dataset mattered for success. This suggests that many expert views give more useful guidance for training AI in medical imaging.
What this means in practice
- •For medical imaging teams: Improve training of AI models for chest X-ray severity assessment by integrating gaze data from multiple radiologists to increase prediction accuracy.
- •For clinical ai developers: Use multi-reader gaze-guided learning to enhance model generalization to new hospital datasets without retraining in some cases.
Authors
Nattakorn Kittisut, Roy M. Gabriel, Marly van Assen, Mohammadreza Zandehshahvar, Carlo N. De Cecco, Elizabeth A. Krupinski, Ali Adibi
Abstract
Radiologists' visual attention can supervise medical image learning, but how reader panel composition affects its value remains untested. We evaluated GazeRIB (Gaze-guided Radiologist-Informed Blackout), which masks unviewed regions during training and uses unmasked images at inference. Eight experts repeatedly read the same 80 chest X-rays (CXRs). We compared individual readers with three-, five- and eight-reader panels across seven architectures and 10 seeds, evaluating 910 models on a separate 2,897-CXR cohort for four-class severity prediction. Every panel improved macro F1 over the no-gaze baseline by 5.33-7.47 percentage points (pp; Bonferroni-adjusted p < 0.007); no individual reader reached significance on this endpoint. Three readers captured 73-80% of the eight-reader gain in macro F1, the area under the receiver-operating-characteristic curve (ROC-AUC) and the area under the precision-recall curve (AU-PRC). On external validation without retraining, all four panels improved ROC-AUC on SIIM-FISABIO-RSNA by +1.93 to +4.24 pp (p <= 0.004); on MIMIC-Eye no panel differed significantly from baseline, whereas the single-reader average declined (-1.40 pp, p = 0.008). Reader composition therefore influenced performance, although transfer depended on the cohort.