Development and Validation of a Deep Learning Strategy for Automated View Classification of Pediatric Focused...

Area of Science:

• Medical imaging and diagnostics
• Artificial intelligence in healthcare
• Pediatric emergency medicine

Background:

• Focused Assessment with Sonography for Trauma (FAST) is crucial for diagnosing hemorrhage in injured children.
• A key limitation of FAST is the variability in clinician expertise for acquiring all necessary ultrasound views.
• Developing automated tools can enhance the consistency and accuracy of FAST examinations.

Purpose of the Study:

• To develop and validate a deep learning-based view classifier for pediatric FAST.
• To assess the accuracy of the classifier on a large, heterogeneous dataset of real-world FAST studies.
• To determine the feasibility of using AI for quality assurance in pediatric FAST.

Main Methods:

• Retrospective cohort analysis of 699 pediatric FAST studies from two emergency departments.
• Development of a deep learning classifier trained on clinician-performed FAST video clips and still frames.
• Dataset split into training (70%), validation (20%), and testing (10%) sets, ensuring patient data was not shared across sets.

Main Results:

• The deep learning classifier achieved an overall accuracy of 97.8% for video clips and 93.4% for still frames.
• High per-view accuracy was observed for still frames: cardiac (96.0%), pleural (99.8%), abdominal upper quadrants (95.2%), and suprapubic (95.9%).
• The classifier demonstrated robust performance across a diverse dataset of 4925 video clips and over 1 million still frames.

Conclusions:

• A deep learning classifier can accurately identify pediatric FAST views, addressing a critical technical limitation.
• Accurate view classification is essential for quality assurance in FAST examinations.
• This AI tool supports the development of advanced deep learning models for improved evaluation of injured children.