Using Machine Learning for Endoscopic Detection of Low‐Grade Subglottic Stenosis: A Proof of Principle

Abstract

AbstractThe current study trains, tests, and evaluates a deep learning algorithm to detect subglottic stenosis (SGS) on endoscopy. A retrospective review of patients undergoing microlaryngoscopy‐bronchoscopy was performed. A pretrained image classifier (Resnet50) was retrained and tested on 159 images of airways taken at the glottis, 106 normal‐sized airways, and 122 with SGS. Data augmentation was performed given the small sample size to prevent overfitting. Overall model accuracy was 73.3% (SD: 3.8). Precision and recall for stenosis were 77.3% (SD: 4.0) and 72.7 (SD: 4.0). F1 score for the detection of stenosis was 0.75 (SD: 0.04). Precision and recall for normal‐sized images were lower at 69% (SD: 4.35) and 74% (SD: 4), with an F1 score of 0.71 (SD: 0.04). This study demonstrates that an image classification algorithm can identify SGS on endoscopic images. Work is needed to improve diagnostic accuracy for eventual deployment of the algorithm into clinical care.