AI trained to detect catheters in pediatric x-rays
Deep learning techniques can be used to detect catheters and tubes in pediatric x-rays, according to a new study published in the Journal of Digital Imaging. These findings could lead to advancements that prioritize x-rays with poorly placed catheters, bringing them to a specialist’s immediate attention.
Lead author Xin Yi, PhD, University of Saskatchewan in Saskatoon, Canada, and colleagues wrote that automated catheter detection has proven to be “a challenging task.”
“Although most catheters have a radiopaque strip to facilitate detection, the strip may become less apparent depending on the projection angle,” the authors wrote. “Catheters may be confused by other similar linear structures like ECG leads and anatomy, including ribs. Additionally, portions of catheters can be occluded by anatomical structures given that radiographs are a 2D projection of a 3D structure.”
The team’s detection system was a five-step process that included data collection, preprocessing, synthetic catheter generation, training and testing. More than 2,500 adult chest x-rays were used for training purposes, chosen at random from a dataset of more than 7,000 examinations. Synthetic catheters were added to the training data. The test dataset, on the other hand, included 35 annotated images from patients younger than four weeks old.
Overall, the team’s neural network was able to detect catheters in less than one second. And when compared with another deep learning approach developed for PICC line tip detection, this new method had a higher precision and recall.
The researchers wrote that their technique should also work with adult x-rays “provided the profile is carefully designed with consideration given to the large variation of catheter and wire types.” They also noted that their efforts could have a significant impact on patient care in the future.
“The approach described in this work may contribute to the development of a system to detect and assess the placement of catheters on x-ray images, thus providing a solution to triage and prioritize x-ray images which have potentially malpositioned catheters for a radiologist’s urgent review,” the authors wrote. “In the future, an automated catheter placement evaluation system may also be used to prepopulate draft radiology reports with text describing the catheters and tubes present on an X-ray image. This may increase the efficiency of radiology reporting.”
Michael has more than 18 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.