Introduction: Although BMI is an international standard measure for obesity in both children and adults, it does not describe the amount and distribution of body fat. In children, it is known that a greater amount of visceral adipose tissue (VAT) compared to abdominal subcutaneous adipose tissue (ASAT) is related to the early onset of diseases, as well as a higher risk of obesity in adulthood 1. Currently there is a widely validated commercial semi-automatic method called AMRA® Researcher, which allows VAT and ASAT to be quantified by studying Dixon MRI sequences 2. However, it is not accessible to everyone due to the financial investment involved. Objective: Our goal was to develop an automatic, simple and free method to quantify ASAT and VAT, with at least the same precision as AMRA® Researcher, this being equal to 0.17 L for VAT and 0.33 L for ASAT, according to the reports obtained from the quantifications made by AMRA Researcher on our own database. Methods: A proprietary database was studied, consisting of in-phase Dixon MRI sequences of 78 Mexican children, aged 7 to 9 years, with different weight profiles (3 were underweight (BMI percentile < 5), 42 were normal weight (BMI percentile 5 - 85), 17 overweight (BMI percentile 85-95) and 16 obese (BMI percentile >95). The subjects were randomly separated into training, validation and testing subsets, containing 42, 18 and 18 subjects respectively. Since the full body scan of each subject was made up of several overlapping volumes, alignment and joining processes was carried out (Fig. 1). Subsequently, other preprocessing techniques was applied on each axial slice of a joined volume, in order to perform an approximate separation of ASAT and VAT. The result of the separation was two 3D volumes that contained approximately only ASAT and VAT respectively (Fig. 2). These two 3D volumes did not necessarily contain all the ASAT and VAT of the subject studied. Afterwards, two (what we call) Total Intensity Maps were created, which were 2D images that resulted from the sum of the intensities of the voxels of the aforementioned 3D volumes (Fig. 3). Two simple Convolutional Neural Networks (CNNs) architectures were proposed, whose inputs were the two Total Intensities Maps of each subject (Fig. 4). These networks were trained, validated and tested with the corresponding subjects. The outputs of the networks were the quantifications of VAT and ASAT respectively. The quantifications made by AMRA Researcher and those made by our methodology were compared using statistical tests, correlations plots and Bland-Altam graphs. Results: Considering the quantifications from testing subjects, there were no significant differences (p > 0.05) between the quantifications made by AMRA Researcher and those made by our methodology. Furthermore, both reached a high correlation. Our methodology obtained a precision for VAT and ASAT equal to 0.17 L and 0.32 L, therefore equaling the precision of AMRA® Researcher (Fig. 5). Conclusions: Although our methodology achieved the precision of AMRA Researcher, the second is a widely validated method. However, our methodology, being automatic, simple, with low demand for computational resources and free, allows its validation by future independent works, after which, it could be useful as an accessible tool that assists specialists in diagnosis, follow-up and prevention of childhood obesity.

Acknowledgements

We would like to thank CONAHCyT for their support of the postdoctoral expenses of Dr. José Gerardo Suárez García.