An ensemble model approach for identifying pulmonary tuberculosis on chest X-ray

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Tuberculosis (TB) is a global infectious disease, and chest x-ray (CXR) examination is a primary screening method. However, manual image interpretation can suffer from low diagnostic efficiency, poor consistency, and limitations imposed by the experience of the interpreters, necessitating the development of intelligent identification tools. This study aims to construct and validate a deep learning ensemble model-based CXR image recognition system for TB, achieving accurate and intelligent screening. The Shenzhen TB CXR public dataset was used as the training dataset, randomly divided into training and test datasets; the Montgomery public dataset was used as an independent external validation dataset. Nine convolutional neural network models (DenseNet201, EfficientNetB7, EfficientNetV2S, InceptionResNetV2, MobileNetV3Large, NASNetLarge, ResNet50V2, VGG19, and Xception) were built using the PyTorch framework and benchmarked. The optimal model was selected using the F1 score as the core metric. Furthermore, using the raw scores of the individual models as input features, four ensemble strategies—linear regression, logistic regression, performance-weighted averaging, and eXtreme Gradient Boosting (XGBoost)—were constructed to evaluate the classification performance and generalization ability of each model. The results show that all nine individual models can distinguish between normal and TB CXR images (all P < 0.001). The VGG19 model performed best on the independent validation dataset, with an area under the curve (AUC) of 0.853 and an F1 score of 0.793. The ensemble models further improved classification performance. The performance-weighted averaging ensemble model achieved an AUC of 0.880 and F1 score of 0.861 on the independent validation dataset, outperforming the optimal individual model and other ensemble strategies. The performance-weighted averaging ensemble model constructed in this study can accurately identify pulmonary TB CXR images and has high generalization performance. It can provide an intelligent tool for primary-level pulmonary TB screening. In the future, it can be further prospectively validated in multi-center, large-sample clinical datasets to promote its clinical translation and application.