Automated Ultrasonographic Detection of Thrombus and Subcutaneous Edema due to Peripheral Intravenous Catheter
Automated system detects thrombi and edema using ultrasonographic images. Thrombi and subcutaneous tissue characteristics were accurately estimated. Machine learning model achieved 0.723 accuracy for thrombus detection. Edema detection had 0.881 accuracy with 0.928 sensitivity. Blood vessel and subcutaneous tissue assessment using ultrasonographic (US) images prevents peripheral intravenous catheter (PIVC) failure but requires training and is often subjective. In this study, we aimed to develop an automated image processing system for detecting thrombi and edema. US images were collected from patients with catheters, featuring subcutaneous thrombi and edema. Using supervised machine learning with fully convolutional networks, we analyzed 263 images for training and 452 images for evaluation. Ground truth data were manually annotated by calculating accuracy, sensitivity, and specificity. In the test dataset of 452 images, 99 thrombi and 359 edema cases were manually detected. In the automatic estimation, thrombi and edema cases were detected in 102 and 360 images, respectively. The accuracy, sensitivity, and specificity were 0.723, 0.383, and 0.818 for thrombus and 0.881, 0.928, and 0.697 for edema, respectively. This study used a new artificial intelligence tool to detect thrombi and subcutaneous edemas in US images. The sensitivity of the thrombus detection was low in this study, and authors of future studies should focus on improving the tool’s performance. This will increase the accuracy and convenience of US imaging for PIVC use.Highlights
Abstract
Aim:
Methods:
Results:
Conclusions:
Overall network architecture of mask R-CNN.
Ultrasonographic image showing edema (blue box) and thrombus (red box).
Ultrasonographic image with auto image processing. The blue box indicates edema, and the red box indicates thrombus.
Contributor Notes