FAIRNESS-AWARE CROSS-DATASET TUBERCULOSIS DETECTION ON CHEST RADIOGRAPHS

Abstract

Our study fills the gap in the literature that does not exist between high within dataset tuberculosis detection and low cross-domain generalization in chest radiography. Our model-driven, two-way (Montgomery ↔ Shenzhen) cross-dataset evaluation is done with an EfficientNet-B3 pipeline where we test architectural improvements and demographic fairness protocols. To the contrary, in contrast to the traditional means, Squeeze-and-Excitation attention and MixStyle regularization produced insignificant or negative gains relative to the baseline. On the other hand, out of domain performance significantly increased with application of fixed, and inverse-frequency gender re-weighting in training, raising the area under the curve, by 0.25 on the Shenzhen-to-Montgomery challenging direction, and demographic confounding was minimized. Analysis of calibration showed large shifts in threshold with dependence on prevalence of classes, demonstrating that deployment-specific operating points need to be used instead of constant clinical cutoffs. Besides, probability reliability increased with temperature scaling, but there remained boundaries-region miscalibration, highlighting the necessity to apply domain-adaptive threshold-calibration when screening limited computational resources. Continual age-probability will point to enduring fairness issues. These results indicate that training-time demographic re-weighting and strict cross-cohort validation has more real-world effect compared to architectural complexity per se. We publish the full training pipeline, refined weights, and predictions to scale up transparent, fair, and reproducible research in medical artificial intelligence.