Automatic quantification of abdominal subcutaneous and visceral adipose tissue in children, through MRI study, using total intensity maps and Convolutional Neural Networks

TL;DR

This study develops a CNN-based method using MRI total intensity maps for automatic quantification of abdominal fat tissues in children, aiming to improve diagnosis and monitoring of childhood obesity-related health risks.

Contribution

The paper introduces a novel CNN approach for automatic VAT and ASAT quantification in children using MRI data, validated against a commercial semi-automatic method.

Findings

High correlation with AMRA R Researcher quantifications

Comparable precision to existing semi-automatic methods

Potential for accessible, free obesity-related disease diagnosis tool

Abstract

Childhood overweight and obesity is one of the main health problems in the world since it is related to the early appearance of different diseases, in addition to being a risk factor for later developing obesity in adulthood with its health and economic consequences. Visceral abdominal tissue (VAT) is strongly related to the development of metabolic and cardiovascular diseases compared to abdominal subcutaneous adipose tissue (ASAT). Therefore, precise and automatic VAT and ASAT quantification methods would allow better diagnosis, monitoring and prevention of diseases caused by obesity at any stage of life. Currently, magnetic resonance imaging is the standard for fat quantification, with Dixon sequences being the most useful. Different semiautomatic and automatic ASAT and VAT quantification methodologies have been proposed. In particular, the semi-automated quantification methodology used commercially through the cloud-based service AMRA R Researcher stands out due to its extensive validation in different studies. In the present work, a database made up of Dixon MRI sequences, obtained from children between 7 and 9 years of age, was studied. Applying a preprocessing to obtain what we call total intensity maps, a convolutional neural network (CNN) was proposed for the automatic quantification of ASAT and VAT. The quantifications obtained from the proposed methodology were compared with quantifications previously made through AMRA R Researcher. For the comparison, correlation analysis, Bland-Altman graphs and non-parametric statistical tests were used. The results indicated a high correlation and similar precisions between the quantifications of this work and those of AMRA R Researcher. The final objective is that the proposed methodology can serve as an accessible and free tool for the diagnosis, monitoring and prevention of diseases related to childhood obesity.