Grade classification of nasal obstruction from endoscopy videos using machine learning

Nasal obstruction (NO), referring to blockage in the nasal cavity, is prevalent, affecting approximately one-third of the adult population. Consequently, diagnosis typically requires a combination of medical imaging techniques and tests, as NO is often subjective. This study aims to automate the grade classification of NO from anterior nasal cavity images using nasal endoscopy as a standalone diagnostic tool for common NO conditions: allergic rhinitis, chronic rhinosinusitis, and deviated nasal septum. To evaluate this, we examined a proposed method based on a support vector machine (SVM) using two explainable features: the number of the middle turbinate (MT) pixels and the MT contact ratio, derived from the segmentation map in our previous publication, involving 73 participants. This evaluation was compared against deep learning methods - ResNet-50 and Vision Transformers (ViT)-tiny using direct images as input. Our SVM-based method achieved an interrater agreement with the manual grade classification of NO, provided by an ear, nose and throat (ENT) consultant, of 0.46 (moderate agreement) and 0.14 (none or slight agreement) on the validation set and testing set, respectively. While the proposed method introduces the first quantitative approach for differentially diagnosing common NO conditions, further investigation into additional features and strategies to obtain video-level grade classification from frame-level classification data is warranted to achieve a suitable interrater agreement for clinical translation. This has the potential to facilitate the transition of ENT examinations from secondary care to primary care settings, consequently reducing unnecessary ENT referrals.Clinical relevance- This study showcases the first successful grade classification of NO using anatomical segmentation maps from the anterior nasal cavity. The findings hold significant clinical potential, aiding in the early detection of NO.