SIIM21: AI flips through abdominal X-rays to spot IVC filters due for removal

A new artificial intelligence network showed promising results for spotting IVC filters on imaging exams and selecting patients eligible for retrieval procedures, according to a study presented Monday during the Society for Imaging Informatics in Medicine Annual Meeting.

IVC filters are intended to be a temporary solution for patients at acute risk of pulmonary embolism. Yet many are left in place far longer than required, opening individuals up to long-term health complications.

Despite these dangers, retrieval rates tend to be low, John Mongan, MD, PhD, associate professor of radiology at the University of California, San Francisco, said during a May 24 virtual session. He noted: “There’s an opportunity to improve patient health by increasing the number of these filters that are retrieved.”

With these factors in mind, Mongan and his UCSF colleagues developed a deep detection network to screen radiographs and find patients with IVC filters, including individuals who may need their implanted device removed.

For their project, the group constructed a new dataset of 5,225 abdominal radiographs from UCSF, including 1,580 with IVC filters. Providers six hours south at UC Irvine created a separate external validation set of 1,424 triple annotated abdominal X-rays, including 573 with filters.

Overall, Mongan said he was pleased with the results. Quantitatively, he noted, they were “excellent.” On the UCSF test set, the tool notched a 96.2% sensitivity and 98.9% sensitivity. And the team was “particularly pleased” with its performance on the external dataset, noting no clinically significant differences.

Mongan did go over a few various errors with the team’s AI detection tool. Those included a handful of cases in which the model didn’t spot IVC filters, including images with obscured filters, low contrast images, and less common filter types against “busy” backgrounds.

“In conclusion, an object detection network can yield highly accurate detection of IVC filters with excellent transferability to other institutions other than where it was trained,” Mongan noted.

Going forward, they hope to screen radiographs daily to spot patients for IVC filter retrieval.

Around the web

Its testing slate will kick off with the diagnostic and interventional radiology subspecialty exams in nuclear imaging on Jan. 10. 

EZaccessMD offers teleconsults with board-certified physicians coupled with mobile diagnostics, hoping to help health plan holders limit costs on facility-based visits. 

A new survey of ortho experts offers key clues to help improve musculoskeletal radiology services and interdisciplinary collaboration. 

Psychology researchers have used machine learning to wring useful two-year dementia trajectory predictions from more than 500 potentially contributing risk factors.

Trimed Popup