Deep Model Projected Statistical Features for Homodyned K-Distribution Parameters Estimation

• Published in: 2023 IEEE International Ultrasonics Symposium (IUS) pp. 1 - 4
• Main Authors: Tehrani, Ali K. Z., Rosado-Mendez, Ivan M., Rivaz, Hassan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 03.09.2023
• Subjects: autoencoder; homodyned K-distribution; Noise reduction; Optimization methods; Parameter estimation; Phantoms; Quantitative ultrasound; Radio frequency; Scattering; tissue characterization; Ultrasonic imaging
• ISSN: 1948-5727
• DOI: 10.1109/IUS51837.2023.10306362

Quantitative ultrasound (QUS) aims to find properties of scatterers which are highly correlated with the tissue microstructure. The homodyned K-distribution is one of the distributions that can model the envelope of RF data under diverse scattering conditions. The parameters of this distribution, the scattering clustering (α) and the ratio of coherent to diffuse scattering amplitude ratio (k) (which we refer to as HK parameters) are considered as valuable QUS parameters for tissue characterization in diagnostic ultrasound. Statistical features from the envelope of the backscattered radiofrequency (RF) data such as point-wise signal-to-noise ratio (SNR), skewness, kurtosis, and the log-based moments have been utilized to estimate HK parameters. Iterative optimization methods or table search can be used to estimate HK parameters from statistical features by comparing the estimated statistical features with theoretical values. In order to obtain HK parameters from the region of interest (ROI), a patch around the sample of interest is selected where the statistical features are calculated. A larger patch size provides more samples for statistical feature calculation, but increases the heterogeneity within the patch which might result in the failure of the optimization method especially for real tissues. Smaller patch size results in deviation of the statistical features from their theoretical values, causing error in the estimated HK parameters. The theoretical values of these statistical features lie in a low dimensional hyperplane since the feasible HK parameters are in low-dimensional manifold and have lower dimension than the statistical features. In this paper, we propose a model projection autoencoder inspired by a denoising autoencoder to project noisy statistical features into the hyperplane of expected values. The reconstructed features can be employed to estimate HK parameters by using any HK parameter estimator.