Ultrasound Scatterer Density Classification Using Convolutional Neural Networks and Patch Statistics.

Area of Science:

• Medical Imaging
• Biophysics
• Computational Biology

Background:

• Quantitative ultrasound (QUS) provides insights into tissue properties like scatterer density.
• Traditional scatterer density classification relies on statistical parameters sensitive to patch size and imaging settings.
• Distinguishing between fully developed speckle (FDS) and underdeveloped speckle (UDS) is crucial for tissue characterization.

Purpose of the Study:

• To adapt and train convolutional neural network (CNN) architectures for quantitative ultrasound (QUS) scatterer density classification.
• To enhance CNN performance by incorporating patch statistics as additional input channels.
• To evaluate the efficacy of CNNs against conventional and deep learning models for scatterer density classification.

Main Methods:

• Adaptation and training of CNN architectures using simulated QUS data.
• Integration of patch statistics as additional input channels to improve CNN performance.
• Implementation of a deep supervision and multitask learning-inspired method to further leverage patch statistics.

Main Results:

• Proposed CNN methods demonstrated superior performance in classifying tissues with varying scatterer densities compared to existing models.
• The CNNs successfully classified scatterer density across different imaging parameters without requiring a reference phantom.
• Incorporating patch statistics significantly improved the accuracy of scatterer density classification.

Conclusions:

• CNNs show significant potential for accurate scatterer density classification in ultrasound imaging.
• The proposed CNN approaches overcome limitations of traditional methods, offering robustness to imaging parameter variations.
• This work highlights the utility of deep learning in quantitative ultrasound for non-invasive tissue characterization.