Hybrid Deep Learning-Based Enhanced Occlusion Segmentation in PICU Patient Monitoring.

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Remote patient monitoring (RPM) utilizes digital technologies and computer vision (CV) as a non-invasive alternative to traditional methods.
• Pediatric Intensive Care Units (PICUs) experience challenges with occlusions that impede accurate image analysis in RPM.

Purpose of the Study:

• To propose a hybrid deep-learning pipeline for effective segmentation of occlusions in PICU RPM.
• To address limited training data scenarios in developing robust CV models for clinical applications.

Main Methods:

• A hybrid segmentation pipeline combining Google DeepLabV3+ and Segment Anything Model (SAM) was developed.
• The pipeline was trained and validated on a small dataset from real-world PICU settings using a Microsoft Kinect camera.
• Performance was evaluated using Intersection-over-Union (IoU) and classification metrics.

Main Results:

• The hybrid model achieved an 85% IoU for occlusion segmentation.
• Classification performance included 92.5% accuracy, 93.8% recall, 90.3% precision, and 92.0% F1-score.
• The proposed method demonstrated an average performance gain of 2.75% over baseline CNN frameworks.

Conclusions:

• The hybrid approach significantly enhances occlusion segmentation for RPM in PICUs.
• This advancement improves the reliability of remote monitoring for pediatric patients.
• The study addresses a critical need for accurate and dependable clinical monitoring solutions.