Deep learning-based phase recognition in anterior nasal endoscopy for handheld devices

Name: Nonpawith Phoommanee

Organization : King Mongkut's University of Technology North Bangkok. Department of Industrial Physics and Medical Instrumentation

Email : nonpawith.p@sci.kmutnb.ac.th

Name: Andrews, Peter J.

Organization : Royal National ENT and Eastman Dental Hospitals. Department of ENT

Email : peterandrews@nhs.net

Name: Leung, Terence S.

Organization : University College London. Department of Medical Physics and Biomedical Engineering

Email : t.leung@ucl.ac.uk

keyword: Medical image analysis

LCSH: Deep learning (Machine learning)

LCSH: Diagnostic imaging

LCSH: Nasal fossa -- Examination

LCSH: Mobile computing

Abstract: Nasal obstruction, affecting approximately one-third of adults, commonly arises from intranasal structures such as the nasal valve, septum, and turbinates. Accurate differential diagnosis depends on systematic visualization of key anatomical phases during anterior nasal endoscopy. However, general practitioners (GPs) in primary care may face difficulties in consistently acquiring and interpreting diagnostically relevant frames, limiting both clinical decision-making and training. To address this, we proposed an automated phase recognition pipeline as a prerequisite for AI-assisted diagnosis. The system identifies four anatomical phases: outside, nostril, anterior, and posterior with the anterior phase being most relevant for diagnosis. We fine-tuned four lightweight pretrained architectures: ResNet-50D, ConvNeXt-v2, DeiT-tiny, and MobileNetv4-hybrid-medium using label smoothing and ordinal penalty loss to enhance classification robustness. A total of 29,973 labeled frames from 54 participants were used in stratified 5-fold cross-validation, with 8,903 frames from 17 participants held out for testing. ConvNeXt-v2 achieved the highest test accuracy (0.775), a quadratic-weighted Cohen’s kappa of 0.774 (indicating substantial agreement with ground truth), and the lowest mean absolute error (0.280). Meanwhile, DeiT-tiny offered the fastest inference speed, with an INT8 quantized latency of 28.95 ms per frame on a single-threaded CPU, suitable for real-time deployment with minimal accuracy trade-off. These findings demonstrate that compact models can generalize effectively to anterior nasal endoscopy and that incorporating ordinal-aware loss aligns with the task’s inherent structure. The proposed system is well-suited for integration into handheld endoscopes, supporting phase-aware ENT assessments and enhancing GP training in primary care.

King Mongkut's University of Technology North Bangkok. Central Library

Address: BANGKOK

Email: library@kmutnb.ac.th

Created: 2025

Modified: 2026-01-07

Issued: 2026-01-07

บทความ/Article

application/pdf

BibliograpyCitation : In IEEE Thailand Section. Antennas and Propagation/Electron Devices/Microwave Theory and Techniques Joint Chapter, Electrical Engineering/Electronics, Computer, Telecommunications, and Information Technology Association, and King Mongkut's University of Technology North Bangkok. The 5th Research, Invention and Innovation Congress (RI2C 2025) (pp.50-53). Bangkok : IEEE Thailand Section

eng

©copyrights King Mongkut's University of Technology North Bangkok

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	RI2C 2025pp.50-53.pdf	1.65 MB

ใช้เวลา

0.038255 วินาที