Key points are not available for this paper at this time.
Introduction Respiratory diseases impose a substantial global clinical burden, and chest radiography remains a widely used first-line imaging modality for evaluating suspected pulmonary abnormalities. Accurate multi-class interpretation is challenging because several pulmonary conditions share overlapping radiographic patterns. Methods We developed PulmoX-Net, a hybrid convolutional neural network that combines the efficient depthwise separable convolutions of Xception with Squeeze-and-Excitation (SE) channel attention to improve feature representation for chest X-ray classification. The model was evaluated on the public Kaggle “X-ray Lung Diseases Images (9 classes)” dataset, which contains 6,743 frontal chest radiographs grouped into nine benchmark-specific, pattern-based categories. Supplementary analyses included image-level stratified split reporting, source-label review, per-class metrics, repeated-run testing, model-complexity profiling, and clinician review of Grad-CAM heatmaps. Results PulmoX-Net achieved an overall test accuracy of 89.23%, precision of 89.92%, recall of 89.38%, F1-score of 88.97%, and macro-average AUC of 0.9842, outperforming the comparator CNN backbones evaluated under the same experimental framework. Discussion These findings suggest that PulmoX-Net is a promising benchmark model for attention-enhanced multi-class chest radiograph classification. However, the results should be interpreted within the limits of a single dataset and its non-standard label scheme; external validation on independent clinical cohorts is required before clinical deployment can be considered.
WU et al. (Fri,) studied this question.