ML Model Classifies, Detects, Segments Structures During NE

by Linda Kossoff •

CLINICAL QUESTION

How can the application of convolutional neural networks (CNNs) facilitate the identification and delineation of structures in the nasal cavity during nasal endoscopy (NE)?

BOTTOM LINE

A CNN-based model can accurately localize and segment turbinates in images obtained during NE, signifying the feasibility of future machine learning (ML) algorithms to interpret NE findings.

BACKGROUND: The heterogeneous nature of nasal cavity structures poses challenges to the interpretation of NE, potentially resulting in clinicians missing subtle signs of sinonasal disease. ML models such as CNNs have shown excellent performance in detecting and segmenting structures in video feeds and might be valuable in NE interpretation.

STUDY DESIGN: Prospective study

SETTING: Department of Otorhinolaryngology, Ochsner Health, New Orleans

SYNOPSIS: Researchers obtained images captured between 2014 and 2023 during NE examinations for the evaluation of sinonasal disease. Eligible images had no distortion, clouding, obstruction, or artifact, and provided clear identification of the surface of the inferior turbinate (IT) and middle turbinate (MT). Researchers randomly selected 2,111 images to manually segment the IT and/or MT. They configured the open-source YOLOv8 object detection model to classify whether a turbinate was present; detect its location; and apply a segmentation mask delineating its borders. The CNN model successfully identified the IT and MT with an average accuracy of 91.5%. The model’s average precision and recall were 92.5% and 93.8%, respectively. At the confidence threshold of 60%, the average F1-score was 93.1%, indicating high segmentation performance. The dataset contained images of labeled turbinates taken from multiple angles with and without comorbid pathology. Authors note this is the first application of CNNs to identify and localize anatomical landmarks on NE images. They also cited their exclusive use of a single center to capture images as a potential source of selection bias and, thus, a limitation of the study.

CITATION: Ganeshan V, Bidwell J, Gyawali D, et al. Enhancing nasal endoscopy: Classification, detection, and segmentation of anatomic landmarks using a convolutional neural network. Int Forum Allergy Rhinol. 2024;14:1521-1524.