Seeking trustworthy automated detection for hip implant

Thursday, November 30 | 9:50 a.m. - 10:00 a.m. | R3-SSIN07-3 | N227B

A deep-learning classifier designed to create a deep-learning “trustworthiness pipeline,” in this case for a total hip arthroplasty (THA) aid in clinical settings, recognized 28 implants on plain hip radiographs, according to recent research out of the Mayo Clinic in Rochester, MN.

According to presenter and research prize fellow Pouria Rouzrokh, MD, MPH, detecting THA implants on plain radiographs is crucial for facilitating revision THA. However, existing deep-learning models lack trustworthiness for clinical deployment. With the goal of improving trustworthiness, Rouzrokh et al trained a deep-learning classifier on 244,248 AP, lateral, and oblique images obtained from 13,375 THA patients from 2000 to 2022.

This trustworthiness pipeline consisted of integrated gradient maps for model explainability; Mondrian cross-conformal prediction (MCCP) for uncertainty quantification that enabled the model to predict more than one label on challenging radiographs; and a framework to detect data outliers.

The team evaluated the pipeline on held-out internal and external test sets and six medical and nonmedical out-of-domain datasets, including ImageNet, RadImageNet, chest x-ray, knee-radiograph, pre-operative hip, and post-operative hip datasets with unseen implants. The group also compared the tool's baseline performance with two board-certified orthopedic surgeons.

According to Rouzrokh, the deep-learning model surpasses all previously reported models in baseline performance and exemplifies a pipeline for making these models trustworthy. The conclusion is based on accuracy F1-scores of 98.9%, 98%, as well as high efficiency scores, according to the researchers.

Further, THA-AID achieved 97.5% accuracy compared to the best surgeon accuracy of 87% on 200 test images. THA-AID also identified 100% of data from out-of-domain datasets, according to Rouzrokh. And eliminating 39.7% of the most outlier external data improved the external set's accuracy, coverage, and efficiency to 100%. 

As a result of this work, Rouzrokh believes THA-AID is a strong candidate for clinical deployment in radiologic and orthopedic surgery applications. Stop by this session for more information.

Page 1 of 2
Next Page