Explainable AI for chest radiographs: sex-stratified fairness auditing in CNN-based pneumonia detection.
Authors
Affiliations (3)
Affiliations (3)
- Department of Diagnostic and Interventional Radiology, Carl Von Ossietzky Universität Oldenburg, Oldenburg, Germany. [email protected].
- IU Internationale Hochschule GmbH, Erfurt, Germany. [email protected].
- IU Internationale Hochschule GmbH, Erfurt, Germany.
Abstract
Artificial intelligence (AI) systems for chest X-ray (CXR) interpretation can achieve high standalone diagnostic accuracy, yet limited transparency raises concerns about subgroup disparities. We audited sex-stratified performance and attention patterns in CNN-based pneumonia detection using the RSNA Pneumonia Detection Challenge dataset (26,684 frontal adult CXRs). ImageNet-pretrained InceptionV3 models were trained across five prespecified random seeds and combined as a seed-ensemble. Analyses used a prespecified, fixed, patient-level sex- and label-balanced test set (n = 1,000; 250 pneumonia-positive and 250 pneumonia-negative per sex) and a single validation-derived operating point held constant across sex. Baseline performance was robust (AUROC 0.859 [0.836, 0.881], Brier 0.155 [0.141, 0.169]). Males showed higher discrimination than females (AUROC 0.882 vs 0.837, gap + 0.045 [+ 0.002, + 0.091]) and a lower false-positive rate (FPR gap - 0.073 [- 0.145, - 0.006]). To contextualize disparities, we quantified Grad-CAM attention relative to lung segmentation masks; in-lung activation was lower in females (60.8% vs 64.7%, M - F + 3.9 pp [+ 1.6, + 6.4]). A combined mitigation (group-balanced sampling + adversarial debiasing) preserved performance (AUROC 0.864 [0.843, 0.885]) while reducing AUROC/specificity gaps and attenuating the false-positive disparity. Quantitative explainability can complement fairness auditing and mitigation in medical imaging.