Failure-aware robustness evaluation of deep learning models for tuberculosis detection under synthetic chest X-ray degradation

Background Deep learning-based systems have demonstrated high performance in automated Tuberculosis (TB) detection from chest X-ray (CXR) images and are increasingly considered for large-scale screening. However, most evaluations rely on high-quality, curated images and do not reflect the degraded imaging conditions frequently encountered in routine clinical practice, particularly in resource-limited settings. This study introduces a failure-aware robustness evaluation of convolutional neural network (CNN) models for TB detection under controlled synthetic CXR degradation scenarios. Methods We fine-tuned three ImageNet-pretrained CNN architectures on the combined Montgomery and Shenzhen datasets (800 CXRs) using stratified patient-level fivefold cross-validation. We applied five degradation types (Gaussian noise, motion blur, JPEG compression, downsampling, contrast reduction) to the held-out test images at three severity levels. We compared discrimination, sensitivity, predictive values, and calibration between clean and degraded conditions. The combined dataset contained 394 TB-positive and 406 non-TB cases (49.25% TB-positive). Results Three CNN models- ResNet-50, DenseNet-121, and MobileNetV2- were evaluated on two public TB CXR datasets comprising approximately 800 images. We systematically applied common image degradations, including noise, blur, compression, reduced contrast, and lower resolution, to the test images. All models performed worse under these conditions. Motion blur produced the largest impact, decreasing sensitivity by up to 21%. Model reliability also declined, as calibration error increased from 0.038–0.051 on clean images to 0.086–0.104 on degraded images. Conclusions The findings show that AI-based TB detection models can be vulnerable to performance degradation under the controlled synthetic imaging conditions evaluated in this study. Robustness and calibration evaluation under degraded inputs should be considered prerequisites for the responsible clinical deployment of AI-assisted TB screening systems, particularly in resource-limited environments.