Machine learning-based dual-energy CT parametric mapping | JoVE Visualize

Area of Science:

Medical Physics
Radiology
Artificial Intelligence in Medicine

Background:

Quantitative parametric maps (effective atomic number, electron density, mean excitation energy, stopping power) are crucial for material identification and radiation dose calculation in CT imaging.
Current physics-based methods for generating these maps from dual-energy CT data have limitations, particularly in accuracy and noise tolerance at low exposure levels.

Purpose of the Study:

To develop and evaluate machine learning (ML) methods for generating quantitative parametric maps from clinical dual-energy CT data.
To compare the performance of ML methods against a clinically-used physics-based reference method across various radiation exposure doses.

Main Methods:

Employed machine learning techniques including historical centroid (HC), random forest (RF), and artificial neural networks (ANN) to learn the relationship between dual-energy CT input and parametric maps.
Trained models using phantoms with known compositions and validated on independent phantom and patient data.
Evaluated precision, accuracy, and computational efficiency against a physics-based reference method at different exposure doses.

Main Results:

ML methods, including RF and ANN, generated more accurate and precise parametric maps compared to the reference method.
Performance advantage of ML methods was most significant at low radiation doses (one-fifth of typical clinical abdomen CT).
Artificial neural networks (ANN) demonstrated excellent accuracy (only 1% less than RF) and superior computational efficiency (15s per map), outperforming RF.

Conclusions:

Machine learning methods significantly outperform the reference method in accuracy and noise tolerance for generating quantitative parametric maps from dual-energy CT.
The ANN method is identified as the most suitable for clinical application due to its optimal combination of accuracy, low-noise performance, and computational speed.
These findings encourage further exploration of ML techniques for advanced quantitative analysis in medical imaging.