Prototype-Driven Hard-Sample Contrastive Learning for Camera-Based Respiratory Imaging Analysis.

Area of Science:

• Medical Imaging
• Machine Learning
• Respiratory Physiology

Background:

• Respiratory spatial patterns are crucial for assessing lung conditions.
• Camera-based analysis with machine learning shows promise but struggles with patient generalization due to data limitations and individual variability.
• Existing data augmentation methods risk distorting essential respiratory pattern information.

Purpose of the Study:

• To develop a novel method, prototype-driven hard-sample contrastive learning (PHCL), for robust camera-based respiratory imaging analysis.
• To address challenges in patient generalization and individual respiratory variability in machine learning models for respiratory pattern analysis.
• To improve the analysis of asymmetric and irregular respiratory activities.

Main Methods:

• PHCL utilizes prototypes and Gini-index distance to classify samples as simple or hard-to-learn.
• It synthesizes new features by blending simple and hard samples to create class transition boundaries.
• Contrastive learning emphasizes feature consistency between prototypes and hard samples across subjects to mitigate individual variability.

Main Results:

• PHCL demonstrated superior performance in extensive experiments conducted in neonatal intensive care and thoracic surgery units.
• The method outperformed traditional image augmentation and advanced feature augmentation techniques.
• PHCL achieved a 1-10% improvement in both accuracy and F1-score compared to existing methods.

Conclusions:

• PHCL offers a significant advancement in analyzing respiratory spatial patterns using camera-based imaging.
• The method effectively mitigates individual respiratory variability and refines class boundaries for improved diagnostic accuracy.
• This work provides valuable insights for analyzing complex respiratory dynamics, particularly asymmetry and irregularity.