Grade classification of nasal obstruction from endoscopy videos using machine learning

Area of Science:

• Otorhinolaryngology
• Medical Imaging
• Artificial Intelligence

Background:

• Nasal obstruction (NO) affects one-third of adults, often requiring subjective diagnosis.
• Current diagnostic methods for NO are complex and time-consuming.
• Automating NO grading could streamline diagnosis and improve patient care.

Purpose of the Study:

• To automate the grade classification of nasal obstruction (NO) from anterior nasal cavity images.
• To evaluate a support vector machine (SVM) based method against deep learning approaches.
• To utilize nasal endoscopy as a standalone diagnostic tool for common NO conditions.

Main Methods:

• A support vector machine (SVM) model was developed using middle turbinate (MT) pixel count and MT contact ratio from segmentation maps.
• The SVM method was compared with deep learning models (ResNet-50, ViT-tiny) using direct endoscopic images.
• The study involved 73 participants, with classification accuracy assessed against an ENT consultant's grading.

Main Results:

• The SVM-based method achieved moderate interrater agreement (0.46) on the validation set and slight agreement (0.14) on the testing set.
• This represents the first quantitative approach for differential diagnosis of common NO conditions.
• Deep learning methods were evaluated but not detailed in the abstract's results.

Conclusions:

• The proposed SVM method shows potential for classifying NO severity using anatomical segmentation maps.
• Further research is needed to enhance interrater agreement for clinical translation.
• This technology could facilitate primary care diagnosis of NO, reducing ENT referrals.