Image-Based Metrics for Ultrasound Speed-of-Sound Estimation and Neural Network Minimization – Roman Denkin (Extra Seminar, OBS the time)

This seminar presents two research directions explored during the first half of my PhD. The first focuses on estimating tissue speed of sound in ultrasound imaging using conventional image analysis metrics, such as correlation, mutual information, and a proposed Tenengrad variant, offering a system-independent alternative to physics-based approaches. The second direction investigates neural network pruning and distillation, demonstrating that large overparameterized fully-connected and convolutional networks can be reduced by up to 25× and 10×, respectively, with minimal performance degradation.

About Roman Denkin