Accurate Lung Segmentation Does Not Improve Deep Learning Detection of Tuberculosis on Chest Radiographs: A Controlled U-Net and ResNet-18 Study

Abstract

Background. Lung field segmentation is routinely placed in front of convolutional neural network (CNN) classifiers for tuberculosis (TB) screening, on the assumption that removing non-lung anatomy helps the classifier concentrate on disease. This assumption has rarely been tested under matched training conditions, at clinically meaningful operating points, or in relation to the quality of the mask itself. This is the research gap the present study addresses.

Methods. We used the 138 posteroanterior chest radiographs of the Montgomery County dataset (80 normal, 58 TB) and their manual lung masks. A four-level U-Net was trained for lung segmentation and compared with a contrast-enhanced Otsu thresholding baseline. An ImageNet-pretrained ResNet-18 was then fine-tuned to detect TB from either the whole radiograph or a lung region of interest (ROI) cropped with the U-Net mask. Both pipelines shared the same stratified five-fold cross-validation, architecture and training schedule, and all metrics were computed on out-of-fold predictions. Areas under the ROC curve (AUC) were compared with DeLong’s test, operating points were judged against the World Health Organization (WHO) minimum accuracy targets for a TB triage test, and changes in predicted probability were related to mask quality, lesion type and Grad-CAM attention.