Deep transfer learning of brain shape morphometry predicts Body Mass Index (BMI) in the UK Biobank

Abstract

Prior studies show that obesity is associated with accelerated brain aging and specific patterns of brain atrophy. Finerscale mapping of the effects of obesity on the brain would help to understand how it promotes or interacts with disease effects, but so far, the influence of the obesity on finer-scale maps of anatomy remains unclear. In this study, we propose a deep transfer learning network based on Optimal Mass Transport (OMTNet) to classify individuals with normal versus overweight/obese body mass index (BMI) using vertex-wise brain shape metrics extracted from structural MRI scans from the UK Biobank study. First, an area-preserving mapping was used to project 3D brain surface meshes onto 2D planar meshes. Vertex-wise maps of brain metrics such as cortical thickness were mapped into 2D planar images for each brain surface extracted from each person’s MRI scan. Second, several popular networks pretrained on the ImageNet database, i.e., VGG19, ResNet152 and DenseNet201, were used for transfer learning of brain shape metrics. We combined all shape metrics and generated a metric ensemble classification, and then combined all three networks and generated a network ensemble classification. The results reveal that transfer learning always outperforms direct learning, and we obtained accuracies of 65.6±0.7% and 62.7±0.7% for transfer and direct learning in the network ensemble classification, respectively. Moreover, surface area and cortical thickness, especially in the left hemisphere, consistently achieved the highest classification accuracies, together with subcortical shape metrics. The findings suggest a significant and classifiable influence of obesity on brain shape. Our proposed OMTNet method may offer a powerful transfer learning framework that can be extended to other vertex-wise brain structural and functional imaging measures.