Purpose or Learning Objective
To assess the validity and reliability of an automated
sarcopenia estimation approach using a deep-learning based ensemble psoas
segmentation and Hounsfield Unit Average Calculation (HUAC) model in comparison
to traditional manual measurements.
Methods or Background
This study retrospectively analyzed 149 CT scans, comparing
sarcopenia assessments between manual HUAC measurements and those derived from
an automated TransUNet-based system. The AI model combined convolutional neural
networks with Transformer blocks to enhance feature extraction and contextual
understanding of muscle tissue, crucial for precise sarcopenia evaluation. The
HUAC was calculated by measuring the area and mean Hounsfield Units (HU) of the
left and right psoas muscles at the L3 vertebra level. Statistical analysis
included mean, standard deviation, correlation, paired t-tests, Bland-Altman
plots, and advanced validation metrics such as Intersection over Union (IoU)
and Dice coefficient to evaluate the model's segmentation accuracy.
Results or Findings
The AI model produced a mean HUAC of 19.66, slightly higher
than the 18.03 from manual assessments, with corresponding standard deviations
of 4.27 and 4.54, respectively. The correlation coefficient of 0.78 indicated
strong agreement between the two methods. The model achieved an IoU of 90% and
a Dice coefficient of 0.90, demonstrating high precision in muscle
segmentation. The systematic bias observed (mean difference of -1.63 HUAC)
highlights areas for further calibration of the AI model.
Conclusion
The integration of AI in sarcopenia assessment through HUAC
calculations offers a promising alternative to manual measurements, providing
speed, reproducibility, and precision. Despite some variance, the AI method
aligns closely with traditional approaches, suggesting that with further
refinement, it could become a standard tool in clinical settings.
