Paper Overview
Field: Computer Vision Authors: Nattakorn Kittisut, Roy M. Gabriel, Marly van Assen, Mohammadreza Zandehshahvar, Carlo N. De Cecco, Elizabeth A. Krupinski, Ali Adibi Published: 2026-10-30 arXiv: 2610.42152
Summary
Radiologists' visual attention can supervise medical image learning, but how reader panel composition affects its value remains untested. The authors 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). The study compared individual readers with three-, five- and eight-reader panels across seven architectures and 10 random seeds, evaluating 910 models on a separate 2,897-CXR cohort for four-class severity prediction.
Key Findings
- Every reader 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 panel's gain in macro F1, ROC-AUC, and AU-PRC.
- In external validation without retraining, all four panels improved ROC-AUC on the SIIM-FISABIO-RSNA dataset by +1.93 to +4.24 pp (p <= 0.004).
- On MIMIC-Eye, all panels showed no significant difference from baseline, while individual readers declined on average (-1.40 pp, p = 0.008).
Conclusion
Reader panel composition affects the value of gaze-guided supervision: even small panels of three readers deliver most of the benefit of larger panels, though transfer performance depends on the target cohort.
Original 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 mac...
--- *Auto-collected on 2026-11-03*