Developing a Deep Learning Approach for Automated Body Composition Prediction in Newborns Using Ultrasound Images.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate measurement of human body composition, including fat mass (FM) and fat-free mass (FFM), is crucial for assessing malnutrition and the efficacy of nutritional interventions.
• Current methods for body composition analysis can be invasive or inaccessible, particularly for vulnerable populations like premature infants.

Purpose of the Study:

• To develop and validate a novel ultrasound scanning protocol integrated with a deep learning pipeline for automated prediction of body composition.
• To identify optimal data processing techniques, suitable deep learning models, and key anatomical locations for predicting FM and FFM.

Main Methods:

• Analysis of a clinical dataset comprising ultrasound images from premature infants (n=65) at the biceps, abdomen, and quadriceps.
• Utilized air displacement plethysmography (ADP) for ground truth FM and FFM measurements.
• Employed pre-processing techniques (denoising, median filtering, data augmentation) and a modified EfficientNet-B1 architecture for automated prediction.

Main Results:

• Pre-processing methods significantly enhanced prediction performance.
• A modified EfficientNet-B1 architecture enabled fully automated body composition prediction from ultrasound images.
• Optimal prediction accuracy was achieved using combined biceps and quadriceps (26.1% MAPE) or biceps, quadriceps, and abdomen (25.32% MAPE) scanning locations.

Conclusions:

• This research presents the first demonstration of deep learning for automated body composition prediction using ultrasound images.
• The developed protocol provides a foundation for a novel, non-invasive method to assess malnutrition in infants.
• Sensitivity analysis indicated that different body parts and tissue thicknesses influence FM and FFM predictions.