GazeRIB：多读者注视引导胸片学习的系统评估
GazeRIB: Systematic Evaluation of Multi-Reader Gaze-Guided Chest X-ray Learning
Nattakorn Kittisut · Roy M. Gabriel · Marly van Assen · Mohammadreza Zandehshahvar · Carlo N. De Cecco · Elizabeth A. Krupinski · et al.
中文摘要
放射科医生的视觉注意力可监督医学图像学习，但读者小组构成如何影响其价值尚未被检验。评估 GazeRIB（注视引导的放射科医生知情黑化），训练时遮蔽未被注视区域，推理时使用未遮蔽图像。八位专家重复阅读同一批 80 张胸片（CXR）。在七种架构和 10 个随机种子上比较单个读者与三、五、八读者小组，在独立的 2,897 张 CXR 队列上评估 910 个模型进行四类严重程度预测。每个小组均使宏 F1 较无注视基线提升 5.33-7.47 个百分点（pp；Bonferroni 校正 p < 0.007）；单个读者在该终点上均未达显著。三读者捕获了八读者在宏 F1、ROC 曲线下面积（ROC-AUC）和精确率-召回率曲线下面积（AU-PRC）上增益的 73-80%。在无需重新训练的外部验证中，四个小组均使 SIIM-FISABIO-RSNA 上的 ROC-AUC 提升 +1.93 至 +4.24 pp（p <= 0.004）；在 MIMIC-Eye 上无小组与基线显著不同，而单读者平均下降（-1.40 pp，p = 0.008）。读者构成因此影响性能，但迁移取决于队列。
关键要点
- 01八位专家重复阅读 80 张胸片，比较单读者与三、五、八读者小组的注视监督效果。
- 02所有小组使宏 F1 较无注视基线提升 5.33-7.47 pp，单读者未达显著。
- 03三读者已捕获八读者在宏 F1、ROC-AUC 和 AU-PRC 上增益的 73-80%。
- 04外部验证中 SIIM-FISABIO-RSNA 上 ROC-AUC 提升 +1.93 至 +4.24 pp，MIMIC-Eye 上无显著差异。
解读
尚无解读。
原始英文摘要
arXiv:2610.42152v1 Announce Type: new 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.