Mitigating Aberration-Induced Noise: A Deep Learning-Based Aberration-to- Aberration Approach

Area of Science:

• Medical Imaging
• Ultrasound Technology
• Artificial Intelligence in Medicine

Background:

• Phase aberration significantly degrades ultrasound image quality by distorting wave propagation through heterogeneous tissues.
• The lack of real-world, non-aberrated ground truth data impedes deep learning model training due to domain shift issues.
• Existing methods struggle with accurate phase aberration correction in practical clinical settings.

Purpose of the Study:

• To develop a deep learning-based method for phase aberration correction in ultrasound imaging that does not require ground truth data.
• To enable direct training of models on real-world experimental data, overcoming the simulation-to-experiment domain gap.
• To improve the robustness and performance of phase aberration correction techniques.

Main Methods:

• A novel deep learning network trained on randomly aberrated radio frequency (RF) data as both input and target output.
• Implementation of an adaptive mixed loss function utilizing both B-mode and RF data for enhanced training efficiency and performance.
• Public release of a large dataset (>180,000 images) with modeled near-field phase screens for training and validation.

Main Results:

• The proposed deep learning method successfully corrects phase aberration without requiring ground truth data.
• The adaptive mixed loss function demonstrated superior convergence and performance compared to conventional methods like mean square error.
• The publicly released dataset and code facilitate further research and development in ultrasound phase aberration correction.

Conclusions:

• Deep learning models can be trained effectively for phase aberration correction using only aberrated data, eliminating the need for ground truth.
• An adaptive mixed loss function is crucial for optimal training of such networks, leading to improved image quality.
• The availability of the dataset and source code will accelerate advancements in ultrasound imaging AI.