From Speech to Sonography: Spectral Networks for Ultrasound Microstructure Classification.

Area of Science:

• Medical imaging
• Ultrasound signal processing
• Machine learning for healthcare

Background:

• Ultrasound radiofrequency (RF) signals contain microstructural tissue information.
• Conventional convolutional neural networks (CNNs) struggle with frequency selectivity in RF data, limiting performance.
• Tissue classification based on microstructural changes is crucial for disease diagnosis.

Purpose of the Study:

• To adapt SincNet for enhanced classification of ultrasound RF data.
• To improve frequency selectivity and model generalization in tissue analysis.
• To develop a more interpretable model for ultrasound-based tissue characterization.

Main Methods:

• Adapted SincNet, a neural network architecture, for processing ultrasound RF data.
• Focused filter learning on low frequency and bandwidth, reducing parameters and improving resolution.
• Introduced Gradient-Weighted Filter Contribution for model interpretability.
• Validated the approach on simulated, phantom, and in vivo (liver steatosis) datasets.

Main Results:

• The modified SincNet demonstrated superior performance in tissue classification across all datasets.
• Achieved better frequency selectivity and reduced model complexity compared to conventional CNNs.
• The interpretability method highlighted important spectral bands for classification.

Conclusions:

• SincNet adaptation offers a promising approach for accurate and interpretable ultrasound-based tissue classification.
• This method can aid in identifying microstructural changes associated with disease.
• The findings suggest potential for improved diagnostic capabilities using ultrasound RF data analysis.