Classification of Catheters and Tubes on Chest Radiographs Using Light-Weight Deep Convolutional Neural Networks

Abstract

Purpose: The research aimed to verify the applicability of low computational complexity and high diagnosis accuracy deep convolutional neural network, using MobileNetV2 to identify the presence of chest catheters and tubes on chest X-ray images. Methods: The dataset of chest X-rays collected from a teaching hospital included the endotracheal tube (ETT), the central venous catheter (CVC), and the nasogastric tube (NGT) datasets. A new method of applying dynamic image size training procedures was implemented and compared with fixed image size training. The idea is to learn more features through dynamic image size training. Transfer learning with pre-trained MobileNetV2 on ImageNet was conducted to accelerate the training process and acquire higher accuracy. Class activation mapping (CAM) was also employed to visualize artificial intelligence (AI) predictions, making AI decisions more explainable. Results: The ETT datasets included 10464 X-ray images, while the CVC and NGT datasets contained 10274 and 9610 images, respectively. The accuracies for ETT, CVC, and NGT are 99.0%, 98.4%, and 96.2% in the validation dataset, while in the testing dataset are 98.8%, 98.6%, and 96.8%, respectively. The area under the receiver operating characteristics (AUROCs) were 0.992, 0.988, and 0.980 in the ETT, CVC, and NGT testing datasets. Conclusion: MobileNetV2 with the dynamic image size achieved dedicated performance in the application of chest catheters and tubes classifications. The similarity of the accuracy between the validation and testing data suggests the good generalization capability of the model.