Leveraging Google AutoML for Pathology Classification: Insights from BAGLS and ContactNBI Datasets in Laryngeal Endoscopy

Introduction: The integration of machine learning (ML) technologies in laryngoscopy has demonstrated significant potential to improve diagnostic accuracy and clinical workflow. However, today’s most effective models rely on advanced computer vision techniques that may be complex and unfamiliar to many laryngologists. Google’s Automated Machine Learning (AutoML) offers an accessible platform to help bridge this technical gap. Our objective was to leverage AutoML to train a model capable of classifying various laryngeal pathologies using well-documented open-source endoscopic images.

Methods: The BAGLS dataset comprises 59,250 videoendoscopy frames (2015-2020) and the ContactNBI includes 11,144 narrow-band images, (2015-2021), both publicly available. Images with incomplete annotations or less than 100 per pathology were excluded. Two single-label classification models were developed using AutoML. The first model was trained on 55,650 images from BAGLS dataset and the second model integrated the BAGLS dataset and 10,023 images from ContactNBI dataset. The models’ performance was evaluated using average precision (AUC-PR), precision, and recall.

Results: The BAGLS model achieved an AUC-PR of 0.807, precision of 77.3%, and recall of 73.6%. The combined dataset achieved an AUC-PR of 0.694, precision of 75.8%, and recall of 59.6%.

Conclusions: The BAGLS-only model achieved higher recall and precision than the combined dataset model. Performance differences can be attributed to dataset heterogeneity (endoscopic images vs. narrow band imaging) and significant class imbalance with healthy images being overrepresented. These findings underscore the accessibility of AutoML, positioning it as an invaluable tool for broader adoption in laryngology.