Deep Learning Model Flags Severe Nasal Structural Abnormalities

A neural network trained on nasal endoscopy videos identifies severe septum-inferior turbinate contact in a retrospective study of 317 patients.

Updated on Nov. 11, 2026 in Asthma

Photograph of a slender steel medical endoscope resting on a pale grey surface under soft daylight, with a plain blurred background.
Researchers reported a deep learning model that flags severe nasal structural abnormalities from endoscopy videos in a retrospective study of 317 patients. AI Illustration. Upload story photo >

Researchers developed a deep learning model that identifies severe nasal cavity structural abnormalities from nasal endoscopy videos. The retrospective study analyzed 628 nasal cavities from 317 patients, with frames labeled by board-certified otolaryngologists.

Why it matters

Nasal structural abnormalities can contribute to breathing difficulties that are often underdiagnosed in primary care settings. An automated tool that flags severe cases could help direct patients toward specialist evaluation.

The Average-Diagnosis strategy reached 85.5% accuracy with 66.2% sensitivity and 90.6% specificity in the bilateral-nasal ResNet34 analysis across 628 cavities from 317 patients. The Severe-Priority strategy traded some specificity for higher sensitivity at 84.6%, and the study did not evaluate symptom severity or functional nasal patency.

The details

Researchers retrospectively analyzed nasal endoscopy videos, with board-certified otolaryngologists assigning frames to severe nasal cavity structural abnormality (S-NCSA), non-severe abnormality, or normal categories. Four ImageNet-pretrained convolutional neural networks were fine-tuned and evaluated using patient-level stratified four-fold cross-validation.

Health Landscape

The study uses specialist otolaryngologist labeling of nasal endoscopy frames as its reference standard, testing whether automated classification can replicate expert judgment on severe structural cases. Its performance figures sit alongside prior endoscopy-based machine learning work that similarly depends on specialist-annotated image sets.

Patients with persistent nasal breathing problems may eventually benefit from automated tools that help primary-care clinicians decide when to refer for specialist evaluation. The model is not yet available for clinical use, and prospective testing is still required before it could change everyday care.

The takeaway

Automated endoscopy analysis may eventually help flag structural nasal problems earlier, but its accuracy must be confirmed in prospective studies before clinical use. Patients with ongoing breathing difficulties should continue to seek specialist assessment.

Further reading

Explore more on Asthma to see how nasal and airway structure relate to breathing conditions.

Source note: This article includes information reported by Nature.