Efficient Pipeline for Rapid Detection of Catheters and Tubes in Chest Radiographs

Catheters are life support devices. Human expertise is often required for the analysis of X-rays in order to achieve the best positioning without misplacement complications. Many hospitals in underprivileged regions around the world lack the sufficient radiology expertise to frequently process X-rays for patients with catheters and tubes. This deficiency may lead to infections, thrombosis, and bleeding due to misplacement of catheters. In the last 2 decades, deep learning has provided solutions to various problems including medical imaging challenges. So instead of depending solely on radiologists to detect catheter/tube misplacement in X-rays, computers could exploit their fast and precise detection capability to notify physicians of a possible complication and aid them identify the cause. Several groups attempted to solve this problem but in the absence of large and rich datasets that include many types of catheters and tubes. In this paper, we utilize the RANZCR-CLiP dataset to train an EfficientNet B1 classification model to classify the presence and placement of 4 types of catheters/tubes. In order to improve our classification results, we used Ben Graham’s preprocessing method to improve image contrast and remove noise. In addition, we convert catheter/tube landmarks to masks and concatenate them to images to provide guidance on the catheter’s/tube’s existence and placement. Finally, EfficientNet B1 reached a ROC AUC of 96.73% and an accuracy of 91.92% on the test set. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.