XComposition: multimodal deep learning model to measure body composition using chest radiographs and clinical data

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Body composition metrics, including visceral fat, subcutaneous fat, and skeletal muscle volume, are crucial indicators for predicting outcomes in cardiovascular disease, diabetes, and cancer.
• Accurate assessment of these metrics is vital for patient prognosis and treatment planning.

Purpose of the Study:

• To investigate the efficacy of deep learning techniques in estimating body composition metrics from standard chest radiographs.
• To develop a multimodal deep learning model that integrates chest X-ray images with basic clinical variables for enhanced body composition analysis.

Main Methods:

• A retrospective cohort of 1118 patients with concurrent abdominal CT scans and chest radiographs was analyzed.
• A multitask, multimodal deep learning model was trained using chest radiographs and clinical data (age, sex, height, weight).
• Abdominal CT-derived body composition served as the reference standard for model training and validation.

Main Results:

• The best-performing multimodal deep learning model demonstrated strong correlations with CT-derived measurements: 0.85 for subcutaneous fat volume, 0.76 for visceral fat volume, and 0.58 for skeletal muscle volume.
• The multimodal approach significantly outperformed unimodal models, particularly for subcutaneous fat volume estimation.
• Mean absolute errors for subcutaneous and visceral fat volumes were 1054 cm³/m² and 667 cm³/m², respectively.

Conclusions:

• A novel multimodal deep learning model effectively estimates body composition using chest radiographs and common clinical variables.
• This AI-driven approach offers a scalable solution for assessing body composition, facilitating large-scale epidemiological and clinical research.
• The findings highlight the potential of leveraging routine chest imaging for non-invasive body composition analysis.