MimickNet

MimickNet replicates clinical-grade ultrasound image post-processing by using deep learning to transform delay-and-sum (DAS) beamformed images to approximate proprietary enhancements such as Dynamic Tissue Contrast Enhanced (DTCE™) used by manufacturers (e.g., Siemens).

Key Features:

• Deep Learning Framework: Transforms conventional delay-and-sum (DAS) beamformed images into images resembling clinical-grade scanner post-processing.
• Unpaired Training: Trains without requiring explicit pairing of post-processed and DAS data, using only post-processed image samples from the target scanner to approximate proprietary post-processing.
• High Structural Similarity (SSIM): Reports SSIM of 0.940 ± 0.018 on a test set of 400 cine-loops, 0.937 ± 0.025 on a prospectively acquired dataset, and 0.928 ± 0.003 on an out-of-distribution cardiac cine-loop after gain adjustment.

Scientific Applications:

• Benchmarking: Provides a baseline for comparing ultrasound image formation and post-processing techniques against clinical-grade standards.
• Model Fine-Tuning: Serves as a pretrained model that can be fine-tuned toward different vendor-specific post-processing methods.

Methodology:

Train a deep learning model to map DAS beamformed images to clinical-grade post-processed images by learning from post-processed image samples from clinical-grade scanners without requiring explicitly paired DAS–post-processed training pairs.