Poster R9 - Toward deep learning-based body composition prediction for newborns using ultrasound images

Introduction: The first 1,000 days of life is a critical period of development, and how well or how poorly children are nourished during this time has a profound impact on their ability to grow, learn and thrive. However, a substantial obstacle is that there is currently no measure that provides accurate/comprehensive assessment of growth and development. The current standard of care uses anthropometric indicators such as weight, height, and body mass index, which are straightforward to acquire, but do not provide a comprehensive assessment of growth and development. More advanced tools, such as air displacement plethysmography which measure body composition measures of fat mass (FM) and fat-free mass (FFM), require expensive equipment and specialized facilities. A promising alternative is portable ultrasound imaging, which given its low cost and ability to evaluate musculoskeletal tissue could act as a feasible assessment tool for nutritional metrics. Our ultimate vision is to apply deep learning interpretation techniques to ultrasound images and develop predictive algorithms for important nutritional and growth metrics, such as FM and FFM. Toward this goal, we present research related to the development of a novel scanning protocol to predict human body composition from ultrasound images.

Materials and

Methods: This research centers on the following questions: (1) Data acquisition: which anatomical locations and image regions are most predictive of FM and FFM? (2) Data pre-processing: which data and processing is most suitable for this application? (3) Data analysis: which baseline deep learning models provide predictions to inform our learning strategy? To evaluate these questions, we conducted secondary analysis on a dataset of 65 premature infants collected following an IRB-approved protocol at the University of Minnesota Medical Center. This dataset consisted of B-mode ultrasound images of infant biceps, abdomen, and quadriceps, and gold standard body composition values (FM and FFM) from air displacement plethysmography. To prepare for analysis, the ultrasound image data underwent a series of standard pre-processing steps. The pre-processed ultrasound images, along with FM and FFM, were utilized for model training and evaluation. The dataset was randomly split into 70% for training, 20% for validation and 10% for testing. The trained model was designed to automatically produce end-to-end predictions of FM and FFM. Experimentation involved ResNet18 and EfficientNet-B1, two optimizers, and two loss functions. These models were chosen for initial analysis since such models are frequently used for regression tasks. Common evaluation metrics, including mean absolute percent error (MAPE), were used to quantify the performance of model regression tasks.

Results, Conclusions, and Discussions: In this study, we demonstrate an end-to-end automated output of ultrasound-based human body composition prediction. When comparing ResNet18 and EfficientNet-B1 architectures, we found that the EfficientNet-B1 architecture with fine tuning and training with MAE loss function resulted in an MAPE value of 25.32%. Subsequently, we conducted a p-value analysis and showed that the first- and second-best configurations were not significantly different, but were significantly better than the others. In summary, the key contributions of this work are: (1) The combination of biceps and quadriceps (BQ) images, as well as the combination of biceps, abdomen, and quadriceps (BAQ), yielded predictions with the lowest MAPE; (2) The implementation of median filtering, denoising, and image augmentation significantly enhanced prediction performance; (3) Among the experimented models, EfficientNet-B1 with fine tuning and trained with MAE loss function demonstrated the most promising baseline performance. Our results provide important guidelines to optimize a novel ultrasound scanning and interpretation protocol to predict human body composition. This has important broader implications for clinical assessment of infant malnutrition.

Acknowledgements (Optional):