Semi-supervised Learning with Contrastive and Topology Losses for Catheter Segmentation and Misplacement Prediction

Tianyu Hwang, Chih-Hung Wang, Holger R Roth, Dong Yang, Can Zhao, Chien-Hua Huang, Weichung Wang

Keywords: Semi-supervised Learning, Contrastive Learning, Topology Loss, Catheter Misplacement, Catheter Segmentation

Abstract: Chest X-ray images are often used to determine the proper placement of catheters, as incorrect placement can lead to severe complications. With the advent of deep learning, computer-aided detection methods have been developed to assist radiologists in identifying catheter misplacement by detecting and highlighting the catheter's path. However, obtaining large, pixel-wise labeled datasets can be challenging due to the labor-intensive nature of annotation. To address this issue, we proposed a novel semi-supervised learning method that combines contrastive loss and topology loss. This method takes advantage of the known topological properties of catheters and does not require extensive labeling. We collected 7,378 chest X-ray images from the National Taiwan University Hospital, which were labeled for misplacement of nasogastric and endotracheal tube catheters, and included pixel-wise annotation. Moreover, the CLiP dataset was used as an unlabeled dataset for semi-supervised learning. We used a hybrid U-Net architecture with an added classification head to perform simultaneous segmentation of the catheter and misplacement classification. Our model achieved average AUC of 0.977 for classification and average Dice score of 0.614 for segmentation.

TL;DR: This study proposed a semi-supervised learning method that combines contrastive loss and topology loss to improve the performance of catheter segmentation and misplacement prediction on chest X-ray images.