Few-Shot Conditional Learning: Automatic and Reliable Device Classification for Medical Test Equipment.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Limited availability of specialized medical image datasets hinders deep learning model training.
• Variability in medical equipment across institutions complicates standardized classification.
• Neonatal resuscitation equipment requires accurate identification, especially in resource-limited settings.

Purpose of the Study:

• To develop a few-shot learning methodology for classifying neonatal resuscitation equipment using a small dataset.
• To integrate a reliability score for enhanced prediction confidence.
• To validate the model's performance in real-world, complex image scenarios.

Main Methods:

• Utilized a pre-trained ResNet with an encoder as a backbone for feature extraction.
• Employed few-shot learning with less than 100 natural images for meta-training.
• Incorporated a reliability score to quantify classification certainty.
• Cross-validated model performance and tested on complex natural images with real-time inference constraints.

Main Results:

• Achieved a median accuracy of over 99% during meta-training with 87 images.
• Demonstrated a lower limit accuracy of 73.4% across different models/folds.
• Observed a median accuracy of 87.25% on complex natural images, with performance impacted by segmentation strategy.
• The model effectively classifies neonatal resuscitation equipment despite data limitations.

Conclusions:

• The proposed few-shot learning methodology is highly effective for classifying specialized medical equipment in data-scarce environments.
• The integration of a reliability score enhances the practical utility of the classification model.
• Future improvements in automatic segmentation can further optimize performance for real-time applications in clinical and simulation settings.