New NBGL framework reduces ultrasound speckle noise while preserving anatomical boundaries

Researchers have introduced NBGL (Noise-Aware Boundary-Enhanced Generative Learning), a novel framework that tackles the longstanding challenge of speckle noise in ultrasound imaging. Speckle noise degrades image quality and obscures anatomical details, complicating diagnosis. Existing reduction methods often over-smooth boundaries or fail to adapt to varying noise levels. NBGL solves this with two coordinated branches: one uses generative learning to suppress speckle, while the other learns boundary-sensitive representations to preserve critical anatomical structures.

The key innovation is the noise-aware interaction weight generation (NIWG) module, which estimates noise levels using 3D Laplacian filtering and a median absolute deviation estimator. This adaptive weight then modulates cross-branch feature coupling via a weighted feature-wise linear modulation (wFiLM) module, enabling robust performance across different noise environments. In extensive tests on 141 3D transvaginal ultrasound volumes with six distinct noise levels, NBGL consistently beat state-of-the-art methods in both speckle reduction and structural preservation, maintaining consistency with annotated anatomical boundaries.