Development of a deep learning model for intussusception using point-of-care ultrasound

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pediatric Emergency Medicine

Background:

• Intussusception is a critical pediatric condition often facing diagnostic delays.
• Point-of-care ultrasound (POCUS) offers a valuable tool for rapid assessment in emergency settings.
• Developing automated diagnostic aids is crucial for timely intervention.

Purpose of the Study:

• To create a novel deep learning model for identifying the characteristic 'target sign' of intussusception in POCUS images.
• To evaluate the efficacy of various machine learning approaches for this specific diagnostic task.

Main Methods:

• A dataset of POCUS images from emergency department visits was curated and preprocessed.
• Machine learning models, including fine-tuning, were trained on static images derived from video clips.
• Model performance was assessed using patient-level, frame-level, and threshold-based analyses.

Main Results:

• Fine-tuning deep learning models outperformed classical, ensemble, and transfer learning methods.
• A threshold-based approach yielded the highest predictive accuracy for intussusception detection.
• The model demonstrated feasibility even with a relatively smaller POCUS image dataset.

Conclusions:

• Deep learning, particularly fine-tuning, is a feasible approach for detecting intussusception on POCUS images.
• This technology holds promise for improving diagnostic accuracy and speed in pediatric emergencies.
• Further development can enhance POCUS utility in resource-limited or time-sensitive scenarios.