Comparison of three generations deep learning models for TB detection based on chest x-rays: CNNs, ViT, and ConvNext
This study aimed to explore the progress of modernized Convolutional Neural Networks (CNNs) in chest X-ray analysis for tuberculosis diagnosis in a real-world setting. In the present study, the diagnostic performance of CNNs (InceptionV3, ResNet50, DenseNet121, EfficientNet-B4, Ensemble learning), a vision transformer (ViT with distillation (DeiT)), and modernized ConvNeXtV2 was compared. Pretrained models were utilized to maintain a consistent baseline using a large, open-source dataset (ImageNet 1K). After fine-tuning on chest x-ray (CXR) studies from tuberculosis (TB) screening, the diagnostic performance of three-generation models was evaluated for binary, label-wise, and multi-class classification. A total of 31,854 CXR images from multiple centers were collected for model development and evaluation. The multi-center showed no significant difference, suggesting that the pretrained models achieved comparable generalization capabilities. However, the label-wise study indicated that ResNet50 had a superior diagnostic performance, achieving the highest area under the curve for all labels, including Exudation (0.97), Proliferation (0.99), Nodules(0.94), Cavitation(0.94), Fibrosis(0.87), Calcification(0.96), Military(0.92), and Pleural Effusion(0.95). Moreover, the ensemble learning combining DenseNet121, InceptionV3, and ResNet50 with weighted voting demonstrated the highest accuracy in 6 out of 8 labels. The multi-class study showed that real-life TB detection was more challenging compared to pneumonia detection. ConvNeXtV2 achieved high performance in distinguishing TB from pneumonia and normal cases, outperforming other models. In conclusion, ConvNeXtV2 demonstrated better performance in TB screening with a relatively small data volume. It could be beneficial for differentiating diseases with subtle differences.