Pixel-Level Tissue Classification for Ultrasound Images.

Area of Science:

• Medical imaging
• Ultrasound technology
• Computational pathology

Background:

• Traditional pixel classification in ultrasound images relies on isolated pixel values and thresholds.
• This method is crucial for carotid plaque composition analysis, impacting disease risk (e.g., strokes) and surgical necessity.
• Threshold-based methods, while dated, remain prevalent in virtual histology applications.

Purpose of the Study:

• To develop an advanced pixel-level tissue classification method for ultrasound images.
• To improve the accuracy and robustness of virtual histology in carotid imaging.
• To introduce novel descriptors that incorporate neighborhood pixel information for enhanced classification.

Main Methods:

• Proposed a classification pipeline involving image normalization, multiscale feature extraction, and machine learning.
• Introduced and evaluated a new pixel descriptor that considers neighborhood information.
• Experimentally assessed various descriptors including statistical moments, texture-based, gradient-based, and local binary patterns.
• Validated the method on a dataset of five tissue types: blood, lipids, muscle, fibrous, and calcium.
• Analyzed the correlation between virtual and real histology with medical specialists.

Main Results:

• The proposed method achieved a classification accuracy of approximately 73% for ultrasound tissue images.
• This represents a significant improvement over state-of-the-art threshold-based methods, which achieved around 54% accuracy.
• Statistical validation confirmed the significance of the accuracy improvement.
• Correlation analysis between virtual and real histology demonstrated the approach's robustness.

Conclusions:

• The novel descriptor-based classification method offers superior accuracy for virtual histology in ultrasound images.
• The approach provides a reliable tool for analyzing carotid plaque composition and assessing disease risk.
• The validated correlation with real histology underscores the clinical potential of this advanced technique.