Objectives: FEES is a standard procedure for diagnosing dysphagia. However, appropriate evaluation of FEES findings is difficult for inexperienced evaluators. Recent progress in deep learning has highlighted the use of artificial intelligence-assisted computer-aided diagnosis (AI-assisted CAD) in medical applications. We investigated the detection accuracy of FEES findings evaluated by inexperienced evaluators with and without the use of CAD.
Methods: The algorithm for FEES-CAD was developed using 25,630 expert-annotated images. A total of 45 inexperienced evaluators from three groups of people (resident doctors, nurses, and medical students), evaluated 32 FEES videos from 32 patients. To confirm the effectiveness of FEES-CAD, first, 32 FEES videos were evaluated without the use of CAD. Second, one half was evaluated with, and one half without, the use of CAD. The detection accuracy of the FEES findings was investigated, and the evaluation results obtained with CAD were statistically compared with those obtained without CAD.
Results: In the first FEES evaluation, the total detection accuracy was 82.2 %. In the second evaluation, the total detection accuracy with CAD was 84.3 %, and that without CAD was 81.7 %. The detection accuracies by the resident doctors, nurses, and medical students with CAD were 90.1 %, 82.6 %, and 79.4 %, respectively, and those without CAD were 83.7 %, 80.9 % and 80.1 %, respectively. In the resident doctors, the detection accuracy was significantly better when CAD was used for evaluation, compared with the non-CAD evaluations.
Conclusion: The present study demonstrated the effectiveness of FEES-CAD in improving the detection accuracy of resident doctors, however, the differences were small.
Keywords: Artificial intelligence-assisted computer-aided diagnosis (AI-assisted CAD); Flexible endoscopic evaluation of swallowing (FEES); Inexperienced evaluator; Swallowing impairment.
Copyright © 2023. Published by Elsevier B.V.