Deep learning screening of otoscopic images for ENT referral: development, validation and deployment of NHREF OtoScan

NHREF OtoScan, a MobileNetV2-based otoscopy referral screening model, achieved 97.0% sensitivity and 75.6% specificity on 718 held-out images for abnormal versus normal ears.

• Why it matters: Non-specialists need a referral decision, yet existing otoscopy deep learning work often targets closed-set diagnosis, uses uncalibrated probabilities, and lacks leakage-aware controls for near-duplicate video-derived datasets.
• What they did: The team merged 4,763 otoscopic images from 16 folders across three public datasets, normalized each view to a fixed illuminated field, split by similarity groups (threshold 0.93), fine-tuned from block9expand, and temperature-scaled probabilities (T = 1.2).
• The result: On the test split, AUROC was 0.974, 15 of 501 abnormal ears were not referred (11 myringosclerosis, none cholesteatoma), and a low-confidence flag (4.0%) separated 87.2% accuracy from 31.0%, supporting calibrated, leakage-controlled deployment with specialist review.

AI summary of the abstract · may contain errors · not medical advice · read the paper.