Pixel-wise body composition prediction with a multi-task conditional generative adversarial network.

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Biomedical Engineering

Background:

• Accurate human body composition analysis is vital for health management and disease prevention.
• Current methods like DXA, CT, and MRI are costly or involve radiation.
• 3D body shape analysis offers a promising, non-invasive alternative for body composition estimation.

Purpose of the Study:

• To develop a novel multi-task deep neural network for pixel-level body composition prediction using 3D body surfaces.
• To accurately estimate subcutaneous and visceral fat distribution from 3D body scans.
• To improve the accuracy and texture of predicted fat maps using an interpreted patch discriminator.

Main Methods:

• Utilized a conditional generative adversarial network (cGAN) for multi-task learning.
• Input: 3D body surface data.
• Output: Pixel-level 2D subcutaneous and visceral fat maps, optimized with an interpreted patch discriminator.

Main Results:

• The method accurately predicts 2D subcutaneous and visceral fat maps from 3D body surfaces.
• Demonstrated superior performance on TCIA and LiTS datasets compared to existing methods.
• Achieved significant improvements: 41.3% for whole body fat, 33.1% for subcutaneous/visceral fat, and 4.1% for regional fat predictions.

Conclusions:

• The proposed deep learning approach provides an accurate and efficient method for body composition analysis.
• This technique offers a non-ionizing, potentially lower-cost alternative to traditional medical imaging for fat assessment.
• The interpreted patch discriminator enhances the textural accuracy of the predicted fat maps.