Deep Learning-Based Body Composition Analysis for Cancer Patients Using Computed Tomographic Imaging.

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Malnutrition affects 30-85% of cancer patients, with current tools missing 20% of at-risk individuals.
• Abnormal body composition, specifically loss of fat and muscle mass, is a key diagnostic criterion for malnutrition.
• Computed tomography (CT) is the gold standard for body composition analysis and is often used in cancer care.

Purpose of the Study:

• To develop a deep learning approach for precise body composition analysis using CT scans in cancer patients.
• To enable early detection of malnutrition by accurately segmenting adipose tissue and skeletal muscle.

Main Methods:

• A deep learning model, Swin UNEt TRansformers (Swin UNETR), was developed to segment adipose tissue and skeletal muscle at the L3 vertebral level.
• The model automatically localizes the L3 vertebra before segmentation.
• The approach utilizes a dataset of 200 abdominal/pelvic CT scans from cancer patients.

Main Results:

• Swin UNETR achieved high segmentation accuracy with Dice scores of 0.92 for adipose tissue and 0.87 for skeletal muscle.
• The model significantly outperformed convolutional neural network benchmarks (2D U-Net) by 2-12% in Dice scores (p<0.033).
• Predictions showed strong agreement with ground-truth data (R² 0.7-0.93), confirming its potential for accurate body composition analysis.

Conclusions:

• The developed deep learning method provides accurate body composition analysis from CT imaging.
• This approach can facilitate earlier malnutrition detection in cancer patients.
• Timely interventions can be supported through improved diagnostic capabilities.