Deep Learning Algorithm Classification of PE Tube Images From a Heterogeneous Pediatric Sample

Saturday, September 28, 2024

10:18 AM – 10:24 AM EDT

Location: MBCC, 225

Disclosure(s):

Sruthi S. Surapaneni, BS:

faculty for this accredited education activity has no relationship(s) with ineligible companies to disclose.

Angela Hong, PA-C:

faculty for this accredited education activity has no relationship(s) with ineligible companies to disclose.

Alexander P. Marston, MD:

faculty for this accredited education activity has no relationship(s) with ineligible companies to disclose.

Courtney A. Hill, MD:

faculty for this accredited education activity has no relationship(s) with ineligible companies to disclose.

Introduction: PE tube checks are necessary but are also burdensome to families and impact access to care for other children seeking otolaryngologic care. Telemedicine care would be ideal but ear exams are limited. This study assessed whether images obtained by an over-the-counter otoscope can be used to train an AI algorithm to recognize whether a tube is in place and patent, is extruded, or is not present in the ear canal or tympanic membrane.

Methods: A prospective study of children aged 10 months to 10 years being seen for PE tube follow-up was carried out in 3 clinics from May to November 2023. A smartphone otoscope was used by non-MDs to take ear exam images and pediatric ENT exam findings were used to label each image as PE tube in place and patent, extruded in canal, or not present. A deep learning image classification algorithm was trained and tested with these images.

Results: A diverse group of 69 children that included neurotypical and divergent and those with craniofacial anomalies yielded 296 images. Multiple colors and types of PE tubes and both wet and dry cerumen were included. Augmentation of the image data yielded 2,072 total images; 1,448 were used to train the algorithm, 311 were used to validate, and 313 were used to test performance of the model after it was trained. The classification accuracy was 97.1%, sensitivity 97.1%, and specificity 98.6%.

Conclusions: A deep learning algorithm was trained with images from a heterogeneous and representative group of children reflecting the general population with PE tubes. It was highly accurate, sensitive, and specific. Its performance was not deterred by the quality of the OTC otoscope image, multiple tube types, cerumen phenotypes, and heterogenous patient cohort. The information gleaned from assessing an image with this algorithm could provide the telemedicine provider valuable adjunctive information for the diverse population of children with PE tubes.

Presenting Author(s)

Sruthi S. Surapaneni, BS

Medical Student
Michigan State University College of Human Medicine
Rochester Hills, MI, United States

Disclosure(s):
faculty for this accredited education activity has no relationship(s) with ineligible companies to disclose.

Co-Author(s)

CB

Corey A. Bryton, BS

Medical Student
Tufts University School of Medicine
Cambridge, MA, United States

Disclosure information not submitted.
NR

Nikhil Rangarajan, MS

Glimpse Diagnostics LLC, United States

Disclosure information not submitted.
AH

Angela Hong, PA-C

Physician Assistant, Pediatric Otolaryngology
The Dr. Elie E Rebeiz Department of Otolaryngology, Tufts Medical Center
Boston, MA, United States

Disclosure(s):
faculty for this accredited education activity has no relationship(s) with ineligible companies to disclose.
Alexander P. Marston, MD

Assistant Profressor
UC Davis Health
Sacramento, CA, United States

Disclosure(s):
faculty for this accredited education activity has no relationship(s) with ineligible companies to disclose.
MV

Mark A. Vecchiotti, MD

Pediatric Otolaryngology
Department of Otolaryngology, Boston Medical Center
boston, MA, United States

Disclosure information not submitted.
Courtney A. Hill, MD

Pediatric Otolaryngologist and Entrepreneur
Self-employed/Glimpse Diagnostics
Minneapolis, MN, United States

Disclosure(s):
faculty for this accredited education activity has no relationship(s) with ineligible companies to disclose.

Senior Author(s)

AS

Andrew R. Scott, MD

Medical Director, Cleft Lip and Palate Team; Pediatric Otolaryngologist
Department of Otolaryngology, Tufts Medical Center
Boston, MA, United States

Disclosure information not submitted.