Variational Ultrasound Channel Data Compression.

Software-defined ultrasound enables high-quality imaging and advanced applications, but real-time streaming of raw channel data over standard digital links requires dedicated compression methods. To address this need, we introduce an end-to-end trainable, nonlinear transform coding model that uses variational inference to map each ultrasound channel data frame into latent variables that can be efficiently encoded using entropy coding. We incorporate a scale hyperprior, transmitted as side information, which adapts the entropy model to each input and enables it to handle the large dynamic range and rapidly varying statistics typical of ultrasound channel data. We train the model using ultrasound channel data frames of human carotid arteries and evaluate the empirical rate-distortion performance at different stages of the processing pipeline using both in vivo and phantom data. Our results show that nonlinear transform coding consistently outperforms traditional methods such as JPEG without degrading reconstructed images or downstream measurements. The scale hyperprior further improves generalizability and coding efficiency without notably increasing computational cost. We achieve compression ratios (CRs) of up to 30:1, potentially enabling channel data rates of digital ultrasound interfaces to match those of standard digital links.