Linking Adolescent Brain MRI to Obesity via Deep Multi-cue Regression Network

摘要

Adolescent obesity has become a significant public health problem for the potential risk of various diseases in later life. Recent biomedical studies have revealed that obesity is associated with structural changes in the brain. Thus the computer-aided analysis of adolescent obesity based on brain MRI is of great clinical value. While previous methods typically rely on hand-crafted MRI features for obesity prediction, we propose to link adolescent obesity and brain MRI through a deep learning framework. The newly released brain MRI data from the large-scale Adolescent Brain Cognitive Development (ABCD) study has paved the way for such an exploration. In this paper, we propose a deep multi-cue regression network (DMRN) for MRI-based analysis of adolescent obesity. Specially, in DMRN, we first design a feature encoding network to automatically extract high-dimensional features from brain MR images, followed by a regression network to predict Body Mass Index (BMI) scores for obesity analysis. To take advantage of other prior knowledge of studied subjects, our DMRN framework further explicitly incorporates the demographic information (e.g., waist circumference) of subjects into the learning process. Experiments have been conducted on 3,779 subjects with T1-weighted MRIs from the ABCD dataset. The results have provided some useful findings: (1) we consolidate the relationship between adolescent obesity and brain MRI as well as demographic information through a deep learning model; (2) we use visualization method to explain the prediction results by highlighting potential biomarkers in the brain MR images that are associated with adolescent obesity.