Comparing fully automated AI body composition biomarkers at differing virtual monoenergetic levels using dual-energy...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Dual-energy CT (DECT) enables virtual monoenergetic imaging (VMI) reconstruction.
• AI algorithms can quantify body composition biomarkers from CT images.
• Understanding VMI's impact on AI biomarkers is crucial for consistent analysis.

Purpose of the Study:

• To evaluate the behavior of AI-driven CT body composition biomarkers across various VMI levels.
• To assess the impact of different VMI energy levels on quantitative biomarkers derived from DECT scans.

Main Methods:

• Retrospective analysis of 88 abdominopelvic DECT scans.
• Reconstruction of images at five VMI levels (40, 55, 70, 85, 100 keV).
• Application of automated AI algorithms for quantifying fat, muscle, bone, calcium, and organ size; analysis of biomarker variation relative to 70 keV.

Main Results:

• Biomarker differences (excluding Agatston score) at 40, 55, 85, 100 keV relative to 70 keV ranged from 39-358, 12-102, 5-48, and 9-75 HU for attenuation-based measures, respectively.
• Area-based biomarkers showed differences of 6-15, 3-4, 2-7, and 0-5 cm²; volume-based biomarkers showed differences of 12-34, 8-68, 12-52, and 1-57 cm³ across the same VMI levels.
• VMI levels below 70 keV demonstrated greater measurement variation compared to levels at or above 70 keV, with Agatston score exhibiting particularly spurious behavior.

Conclusions:

• Automated AI CT biomarkers demonstrate stability at VMI levels of 70 keV and higher.
• Lower VMI energy levels (<70 keV) result in significant deviations and should be avoided for reliable body composition assessment.
• The findings support the use of 70 keV VMI as a consistent reference for AI-based body composition analysis with DECT.