Expansion of neck adipose tissue (NAT) is an understudied trait in obesity biology. NAT can be quantified using dual X-ray absorptiometry (DXA), a tool commonly used in body composition studies. However, the neck region is not defined in standard regional outputs from a body composition DXA acquisition; instead, quantifying NAT relies on a scanner-dependent software, where manual input is required to define a non-standard region of interest. This makes specialised body composition studies at scale very time-consuming. We thus developed an automated pipeline for NAT estimation from DXA using convolutional neural networks. We investigated whether predicting measurements with a prior step of cropping a region of interest, using automatic landmark prediction, was better than directly predicting from the entire image. We then compared our proposed architecture to the ResNet50 architecture, a well known model used as a basis for transfer learning experiments in many tasks including classification and regression. For the direct and the two-step prediction, both models displayed high performance accuracy. The anatomical landmark placement model performed within three pixels accuracy, and NAT estimation errors for both models were within less than 2.5% points of mean absolute error. However, the proposed model outperformed ResNet50 in the direct prediction, where ours had a mean absolute error of 2.42 against 5.58 for ResNet50. To ensure that the direct predictions are specifically focusing on neck fat, and that the results did not arise from chance correlations with other adipose tissue depots, we generated activation maps which highlighted that the network is focusing on the neck region. This work will enable large-scale population analyses of non-standard regions of interests across a variety of DXA analyses.
3 - 12