Distilling knowledge for phase aberration correction using deep neural networks

Phase aberration is a major source of image degradation in ultrasound, caused by spatial inhomogeneities in the speed of sound within tissue. While traditional correction techniques and recent deep learning approaches have shown promise, most rely solely on aberrated data during training to predict the aberration profile or sound speed map. This approach overlooks potentially informative non-aberrated counterparts, even when such data is available in controlled settings. To address this gap, we propose, for the first time, a novel teacher-student knowledge distillation framework that leverages privileged supervision during training to improve aberration correction. Knowledge is transferred from teacher to student via softened output distributions and spatial attention transfer, consistent with the learning using privileged information paradigm. Both networks share a MobileNetV2 backbone modified for regression. In addition, we fabricate custom-made physical aberrating layers to acquire experimentally aberrated ultrasound images and evaluate the method on real data. Quantitative analysis using contrast and generalized contrast-to-noise ratio, as well as strength and correlation length of residual aberration profiles, demonstrates that the proposed method consistently outperforms both delay-and-sum beamforming and the baseline method trained without distillation. We have released the corresponding datasets on http://data.sonography.ai/ to support benchmarking and reproducible research.