Comparative study of Deep Learning models for segmenting abdominal adipose tissue in CT scans

Abstract

Purpose: Body composition analysis is a test that measures the proportion of various tissues of a person’s body. It serves as an indicator for certain medical conditions such as metabolic syndrome, cancer, diabetes, or cardiovascular disease. Traditionally, these analyses are done using
anthropometric methods or clinical tools that provide an approximated result. Using the family of U-NET Deep Learning architectures, we perform a fully automatic segmentation of visceral and subcutaneous abdominal adipose tissues. We study these segmentation results and compare them against semiautomatic and manual generated ground truths. Materials and methods: We employ several variations of the U-Net Deep Learning architecture: U-Net, R2U-Net, Attention U-Net, and
Attention R2U-Net. These methods were trained on a dataset, which consists of 554 images from the
Hospital Universitario San Ignacio and IDIME Institute in Bogota, Colombia, collected from 2015 to 2017. This dataset contains annotations for three different tissues: visceral fat, subcutaneous fat and other tissue generated through semiautomatic segmentation tools. Results: Sørensen-Dice index is used as the evaluation metric against the ground truth which consists of manual segmentations performed by experts. We obtained that the U-Net architecture was the most accurate in terms of overall body composition segmentation, with a mean Dice score of 93.0%, followed closely by the Attention U-Net architecture. Conclusions: We found that the U-Net and Attention U-Net architectures are more suited for body composition analysis. The segmentation results produced by these methods could be used to obtain precise metrics and help physicians understand the patient’s physical condition.