CompositIA

Computation of body composition scores from toraco-abdominal CT scans

Abstract

Body composition scores provide a quantitative assessment of the volume and physical properties of specific tissues within the body. Their usage is particularly relevant in clinical and research settings to evaluate nutritional status, risk of certain diseases, and monitoring changes in body composition over time or in response to therapies.
Standard scores include subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT), skeletal muscle area (SMA) at the level of third lumbar vertebra (L3), and density of the spongiosa part of the first lumbar vertebra (L1). These scores are typically derived from thoraco-abdominal computed tomography (CT) scans. However, their computation involves manual operations such as slice selection and area segmentation, which are time-consuming and prone to human bias.
Here, we propose CompositIA, a pipeline to automate the computation of seven CT-based body composition indices based on artificial intelligence techniques. The pipeline consists of three main steps: automatic identification of the L1 and L3 vertebrae, segmentation of image slices at the L1/3 spinal level, and quantification of body composition indices. The system was trained and cross-validated using a k-fold strategy on 205 CT scans. Moreover, it was validated on an independent dataset encompassing 54 CT scans. Results indicate a strong positive linear relationship and good agreement between the automatically computed scores, and ground truth scores manually computed by a pool of radiologists. This was confirmed by regression analyses, Bland-Altman analysis, and resulted in mean percentage relative errors below 15% (accuracy > 85% in detecting CT slices intersecting the L1 and L3 vertebrae, volumetric Dice coefficient > 0.85 compared to manually segmented CT scans). CompositIA is made available as an open source software package for research purposes.