PMDA-GAN: Physical-aware Multi-Domain Adaptive Generative Adversarial Network

Abstract

Ultrasound imaging is a core clinical modality, but its quality is severely degraded by inherent multiplicative speckle noise. Traditional deep learning-based denoising methods struggle in real-world scenarios due to the lack of paired clean images and speckle statistical variations across distinct pathological categories. To address this, we propose PMDA-GAN, a framework that requires no clean references. It leverages the statistical independence of speckle between adjacent frames in clinical videos for noise-to-noise self-supervised learning. The generator combines a shared physical encoder with LoRA adapters, using instance normalization and shallow feature fusion to preserve texture and structural details. The discriminator employs pathology-aware domain-conditional embeddings and a PatchGAN adversarial constraint, coupled with a feature matching loss, to align texture distributions and suppress noise amplification. Evaluated on diverse clinical ultrasound sequences covering various pathological types, PMDA-GAN significantly improves NRR without paired clean images, outperforming existing unsupervised and self-supervised methods in multi-domain denoising.