Training of a novel artificial intelligence algorithm on the first online database of laryngeal vessels of the vocal folds using contact endoscopy and narrow band imaging

Introduction Intraoperative images of vocal fold vessels using contact endoscopy and narrow band imaging (CE-NBI) have already been successfully used for endoscopic differentiation between benign and malignant vocal fold lesions and for training artificial intelligence (AI) algorithms. The first online database of such CE-NBI images was published in 2022 to promote cooperation between laryngological centers and the further development of AI-based approaches.

Material and methods The online database contains 11,144 CE-NBI images from 210 patients with histologically proven benign and (pre)malignant vocal fold lesions. In the present study, 80% of these images were used for training and 20% for testing a novel AI-based (Convolutional Neural Network-CNN) approach to differentiate between benign and malignant laryngeal lesions. Finally, the sensitivity, specificity and accuracy of the method in the automated classification of the test images were calculated.

Results The developed algorithm was trained with the CNN-based AI approach using 8,915 CE-NBI images from the online database. Applied to the 2,229 test images, a sensitivity of 82.2%, a specificity of 90.2% and an accuracy of 87.8% could be reached.

Conclusion The results of the presented AI-based approach regarding the diagnostic quality of the method are comparable to previously published studies on the manual or automated evaluation of CE-NBI images. The online database is a valuable tool for the further development of AI algorithms in the diagnosis of vocal fold lesions.

Publication History

Article published online:
12 May 2023