Automatic contrast phase estimation in CT volumes.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Accurate contrast phase identification is crucial for medical image analysis.
• Manual phase labeling can be time-consuming and subjective.
• Automating this process can enhance efficiency and consistency.

Purpose of the Study:

• To develop and validate an automatic algorithm for contrast phase labeling in medical imaging.
• To improve the accuracy and speed of identifying different contrast phases.
• To create a robust system that handles potential failures in region detection or imbalanced datasets.

Main Methods:

• An automatic algorithm was proposed for contrast phase labeling based on intensity changes in anatomical regions (aorta, vena cava, liver, kidneys).
• Regions were detected using a learning-based discriminative algorithm.
• Multi-class LogitBoost classifiers were employed for independent phase estimation within each region, forming a decision tree for final labeling.

Main Results:

• The algorithm demonstrated high accuracy across different phases: 96.2% for native phase, 92.2% for hepatic dominant phase, 96.7% for hepatic venous phase, and 86.4% for equilibrium phase.
• The system achieved an average processing time of 7 seconds per volume on a dataset of 1016 volumes.
• The decision tree approach combining multiple regions proved advantageous for handling detector failures and imbalanced training data.

Conclusions:

• The proposed automatic algorithm provides accurate and efficient contrast phase labeling.
• The system's robustness is enhanced by combining information from multiple anatomical regions.
• This automated approach has the potential to significantly streamline radiological workflow and improve diagnostic accuracy.