Related Experiment Video
Updated: Mar 31, 2026
Robot-Assisted Transcanal Endoscopic Ear Surgery for Congenital Cholesteatoma
Published on: December 15, 2023
Deep learning-assisted otoscopic screening for paediatric otitis media: feasibility of edge deployment
Changwei Lv1, Desheng Jia2, Bo Gao1
1School of Sino-German Robotics, Shenzhen University of Information Technology, Shenzhen, China.
Insights
A deep learning model aids in screening pediatric otitis media from otoscopic images. INT8 quantization maintains high accuracy while enabling efficient point-of-care device integration.
Area of Science:
- Medical imaging
- Artificial intelligence
- Pediatrics
Background:
- Otitis media is a common childhood illness.
- Distinguishing between acute otitis media (AOM), otitis media with effusion (OME), and normal tympanic membranes via otoscopy is challenging in clinical settings.
- There is a need for decision support tools for screening and pre-diagnostic triage of pediatric otitis media.
Purpose of the Study:
- To develop and validate a deep learning model for otoscopic screening of pediatric otitis media.
- To assess the feasibility of using this model for embedded point-of-care applications.
Main Methods:
- A dataset of 19,522 retrospective otoscopic images was collected and labeled as AOM, OME, or normal.
- A MobileNetV3-Small classifier was trained and evaluated on a balanced test set.
- Three post-training variants (float32, dynamic-range INT8, full-integer INT8) were compared for STM32H7 embedded deployment.
Main Results:
- The deep learning model achieved high accuracy across all variants: 97.33% (float32), 97.33% (dynamic-range INT8), and 97.17% (full-integer INT8).
- Sensitivity ranged from 96.00% to 98.00%, and specificity ranged from 98.00% to 99.75%.
- Full-integer INT8 quantization significantly reduced model size (11.52 to 2.91 MiB) and memory buffer requirements (918.22 to 243.20 KiB).
Conclusions:
- Deep learning-assisted otoscopic screening shows promise for supporting clinical triage of pediatric otitis media.
- INT8 quantization effectively preserves diagnostic accuracy while enhancing feasibility for embedded systems.
- Further prospective, multicenter validation is recommended.
Background:
Otitis media is common in children. Otoscopic differentiation of acute otitis media (AOM), otitis media with effusion (OME) and normal tympanic membranes is often challenging in busy clinics and telemedicine, supporting the need for decision support in screening and pre-diagnostic triage.
Objective:
To develop and validate a deep learning model for otoscopic screening of paediatric otitis media and assess feasibility for embedded point-of-care use.
Methods:
We retrospectively collected 19,522 otoscopic images (Shenzhen Children's Hospital, 2016-2019) from multiple otoendoscopic systems and labelled AOM, OME or normal using recorded diagnoses and otoscopic findings. A MobileNetV3-Small classifier was trained on a development set and evaluated once on a locked, balanced test set (n = 600; 200/class). Three post-training variants were compared in an STM32H7-targeted workflow: float32, dynamic-range INT8 and full-integer INT8.
Results:
Accuracy was 97.33% (float32), 97.33% (dynamic-range INT8) and 97.17% (full-integer INT8). Sensitivity and specificity were 96.00-98.00% and 98.00-99.75%, respectively. Full-integer INT8 reduced weight storage from 11.52 to 2.91 MiB and activation buffer from 918.22 to 243.20 KiB.
Conclusions And Significance:
Deep learning-assisted otoscopic screening may support clinician-facing triage for paediatric otitis media; INT8 quantisation preserved accuracy while improving embedded feasibility. Prospective multicentre validation is warranted.
