A2A Framework Boosts Ultrasound SNR by 84.8% Without Pre-training

Ultrasound images suffer from electronic and speckle noise that complicates clinical interpretation. Traditional denoising relies on explicit noise models, while learning-based methods require massive labeled data and fail under domain shifts. In a new paper on arXiv, Jiajing Zhang and colleagues introduce A2A, a pure test-time training framework for one-shot denoising of synthetic aperture ultrasound (SAU). The model extracts anatomical structure from shuffled sub-apertures using self-contrastive learning in pyramid latent spaces, then discards the noise component. Because A2A trains on the actual noisy sample at inference time, it inherently avoids domain mismatch and pre-training costs.

In simulation experiments with electronic noise from 0 to 30 dB and various inclusion geometries, A2A delivered a 69.3% SNR improvement and 34.4% CNR improvement. In vivo results across six echocardiographic views, liver, and kidney scans—using just two aperture signals—showed even greater gains: 84.8% SNR and 25.7% CNR. The framework produces clear images across diverse targets and configurations, paving the way for more reliable anatomical visualization and functional assessment without the burden of large labeled datasets or pre-trained models.