The Impact of Deep Learning on Determining the Necessity of Bronchoscopy in Pediatric Foreign Body Aspiration: Can Negative Bronchoscopy Rates Be Reduced?

Abstract

Introduction

This study aimed to evaluate the role of deep learning methods in diagnosing foreign body aspiration (FBA) to reduce the frequency of negative bronchoscopy and minimize potential complications.

Methods

We retrospectively analysed data and radiographs from 47 pediatric patients who presented to our hospital with suspected FBA between 2019 and 2023. A control group of 63 healthy children provided a total of 110 PA CXR images, which were analysed using both convolutional neural network (CNN)-based deep learning methods and multiple logistic regression (MLR).

Results

CNN-deep learning method correctly predicted 16 out of 17 bronchoscopy-positive images, while the MLR model correctly predicted 13. The CNN method misclassified one positive image as negative and two negative images as positive. The MLR model misclassified four positive images as negative and two negative images as positive. The sensitivity of the CNN predictor was 94.1 %, specificity was 97.8 %, accuracy was 97.3 %, and the F1 score was 0.914. The sensitivity of the MLR predictor was 76.5 %, specificity was 97.8 %, accuracy was 94.5 %, and the F1 score was 0.812.

Conclusion

The CNN-deep learning method demonstrated high accuracy in determining the necessity for bronchoscopy in children with suspected FBA, significantly reducing the rate of negative bronchoscopies. This reduction may contribute to fewer unnecessary bronchoscopy procedures and complications. However, considering the risk of missing a positive case, this method should be used in conjunction with clinical evaluations. To overcome the limitations of our study, future research with larger multi-center datasets is needed to validate and enhance the findings.

Type of study

Original article.

Level of evidence

III.