CEUS-MSSDM: A Multi-Stage Self-Supervised Diffusion Model for Thyroid CEUS Denoising

Abstract

Contrast-enhanced ultrasound (CEUS) provides vital thyroid nodule diagnostics, yet inherent image noise necessitates advanced denoising. Supervised methods are constrained by scarce clean references, while Noise2Noise (N2N) fails on CEUS due to unmatched noise pairs and signal-dependent noise violations. We propose CEUS-MSSDM: a multi-stage diffusion framework integrating self-supervised learning with hierarchical noise modeling. Our solution combines statistical noise characterization, Markov chain state matching for temporal consistency, and conditional diffusion generation for microstructure preservation. Extensive T-CEUS experiments demonstrate CEUS-MSSDM’s superiority, achieving up to 0.31 SNR and 0.10 CNR gains over state-of-the-art methods while enhancing diagnostic quality.

Ähnliche Arbeiten