Non-inferiority acceptance testing for a CT body composition algorithm

Daniel Blezek, Michael Moynagh, Hillary Garner, Patrick Navin, Doris Wenger

Keywords: body composition, non-inferiority testing, CT, segmentation

Abstract: Value of an algorithm in a clinical setting is difficult to gauge. Our hypothesis for this study is that the acceptance rate of an automated segmentation algorithm for determination of body composition is non-inferior to manual segmentation. A panel of four abdominal radiologists reviewed blinded segmentation results from human raters and the algorithm and were asked to accept or reject results based on the question "I am confident these results are reasonably accurate, and would record the measurements into the patient record". Segmentations were accepted at a rate of $82\%$ overall, $82\%$ for the algorithm and $84\%$ for manual and the algorithm was non-inferior to manual segmentation with $p < 0.025$, however, the study did not reach $80\%$ power. We conclude the algorithm performs as well as manual segmentation but a larger study is required for proper power.

Paper Type: validation/application paper

Primary Subject Area: Validation Study

Secondary Subject Area: Application: Radiology

Paper Status: original work, not submitted yet

Source Code Url: This paper is a validation a published AI algorithm and does not have source code to release.

Data Set Url: Patient data has not been approved for public release.

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.