Improving Preterm Infants' Joint Detection in Depth Images Via Dense Convolutional Neural Networks

Area of Science:

• Medical Technology
• Artificial Intelligence
• Neonatal Care

Background:

• Preterm infant spontaneous motility is a key indicator for motor and cognitive impairments.
• Current movement assessment relies heavily on subjective visual inspection by clinicians.

Purpose of the Study:

• To develop a 2D dense convolutional neural network (denseCNN) for detecting preterm infant joints in depth images.
• To enhance the accuracy of joint and connection detection compared to previous models.

Main Methods:

• Utilized a 2D dense convolutional neural network (denseCNN) architecture.
• Trained and tested the model on depth images acquired in neonatal intensive care units.
• Evaluated model performance on a mid-range laptop for real-time processing.

Main Results:

• Achieved a median recall value of 0.839 for joint detection.
• Demonstrated real-time prediction capability at 0.014 seconds per image.
• Confirmed the model's effectiveness even in resource-constrained environments.

Conclusions:

• The denseCNN model significantly improves the objective assessment of preterm infant motor function.
• Real-time performance enables potential integration into clinical and domestic monitoring systems.
• Facilitates early detection and intervention for motor and cognitive impairments in preterm infants.