Objectives/hypothesis: Create an autonomous computational system to classify endoscopy findings.
Study design: Computational analysis of vocal fold images at an academic, tertiary-care laryngology practice.
Methods: A series of normal and abnormal vocal fold images were obtained from the image database of an academic tertiary care laryngology practice. The benign images included normals, nodules, papilloma, polyps, and webs. A separate set of carcinoma and leukoplakia images comprised a single malignant-premalignant class. All images were classified with their existing labels. Images were randomly withheld from each class for testing. The remaining images were used to train and validate a neural network for classifying vocal fold lesions. Two classifiers were developed. A multiclass system classified the five categories of benign lesions. A separate analysis was performed using a binary classifier trained to distinguish malignant-premalignant from benign lesions.
Results: Precision ranged from 71.7% (polyps) to 89.7% (papilloma), and recall ranged from 70.0% (papilloma) to 88.0% (nodules) for the benign classifier. Overall accuracy for the benign classifier was 80.8%. The binary classifier correctly identified 92.0% of the malignant-premalignant lesions with an overall accuracy of 93.0%.
Conclusions: Autonomous classification of endoscopic images with artificial intelligence technology is possible. Better network implementations and larger datasets will continue to improve classifier accuracy. A clinically useful optical cancer screening system may require a multimodality approach that incorporates nonvisual spectra.
Level of evidence: NA Laryngoscope, 2020.
Keywords: Artificial intelligence; endoscopy; machine learning; neural network; optical biopsy; vocal fold lesions.
© 2020 The American Laryngological, Rhinological and Otological Society, Inc.