Speckle2Self: Self-Supervised Ultrasound Speckle Reduction Without Clean Data

Xuesong Li, Nassir Navab, Zhongliang Jiang

Abstract

Medical ultrasound imaging is frequently degraded by speckle noise, a deterministic interference pattern that obscures anatomical detail. Traditional deep learning denoising methods are often inapplicable due to the lack of "clean" ground-truth images and the spatially correlated nature of speckle. We present Speckle2Self, a novel self-supervised framework that utilizes Multi-Scale Perturbation (MSP) to separate underlying tissue structures from noise using only single noisy scans. Our approach outperforms existing self-supervised and traditional denoising methods, maintains real-time inference speeds, and demonstrates robust generalization across different ultrasound hardware.

Introduction

Medical ultrasound (US) is a cornerstone of diagnostic imaging due to its non-invasive, real-time, and cost-effective nature. However, its utility is often limited by speckle noise—a grainy, multiplicative pattern resulting from the coherent interference of backscattered acoustic waves. Unlike additive white Gaussian noise found in natural images, US speckle is deterministic and tissue-dependent, making it particularly challenging to remove.

Standard supervised denoising requires pairs of noisy and clean images, which are impossible to acquire in clinical ultrasound. Furthermore, existing self-supervised methods like Noise2Noise or Noise2Void fail because they assume noise is either random across multiple acquisitions or spatially independent at the pixel level. Speckle2Self addresses these limitations by introducing a perturbation strategy that breaks the deterministic link between tissue and speckle.

Methodology: Multi-Scale Perturbation

The core innovation of Speckle2Self is the Multi-Scale Perturbation (MSP) strategy. Since a single tissue structure produces a consistent speckle pattern at a fixed scale, the authors propose artificially varying the scale to create the diversity needed for self-supervision.

The Speckle2Self training and inference pipeline, highlighting the Multi-Scale Perturbation (MSP) strategy and dual loss functions.

The framework operates as follows:

  1. Perturbation: A single noisy input image is downsampled and then upsampled at multiple scales (e.g., 1.0x, 0.5x, 0.25x). This process preserves the global anatomical structure but distorts the high-frequency speckle patterns differently at each scale.
  2. Architecture: The model employs a multi-encoder, shared-decoder architecture. Each scale-specific encoder ($E_S, E_M, E_L$) processes a differently perturbed version of the image, feeding into a common decoder ($D$).
  3. Dual Loss Optimization: The network is trained using a combination of two loss functions:
  4. Reconstruction Loss ($\mathcal{L}_{rec}$): A Mean Squared Error (MSE) loss that ensures the output remains faithful to the input's structural content.
  5. Consistency Loss ($\mathcal{L}_{con}$): An L1 loss that forces the outputs from different scales to be identical. Because the anatomy is invariant across scales while the speckle is not, the network learns to suppress the scale-variant noise and retain the scale-invariant tissue structure.

Experimental Setup and Datasets

The performance of Speckle2Self was validated using both synthetic simulations and real-world clinical data:

  • Synthetic Data (S-I to S-III): Generated via the Matlab Ultrasound Toolbox (MUST), these datasets provide rare ground-truth "clean" targets for quantitative evaluation using metrics like SSIM and LPIPS.
  • In-vivo Carotid Images: Real B-mode images were collected using a Clarius L7 portable scanner for training (419 frames) and testing (104 frames).
  • Cross-Device Evaluation: To test generalization, additional datasets were acquired using Siemens Juniper and Cephasonics machines.

Results and Discussion

Speckle2Self demonstrates a significant advancement over both traditional mathematical filters (such as BM3D and SRAD) and previous self-supervised AI models. On synthetic benchmarks, it achieved superior structural preservation. On clinical data, it produced the highest homogeneity scores, indicating effective smoothing of noise without sacrificing edge definition.

Qualitative comparison of Speckle2Self against built-in denoising across three different ultrasound systems, demonstrating robust cross-device generalization.

Key findings include:

  • Generalization: A model trained exclusively on Clarius scanner data successfully denoised images from Siemens and Cephasonics systems without prior exposure, a feat facilitated by the universal physics of speckle formation.
  • Efficiency: With an inference time of approximately 7ms, the model is suitable for integration into real-time ultrasound systems operating at 30–60 Hz.
  • Downstream Utility: Pre-processing with Speckle2Self improved the performance of the Segment Anything 2 (SAM2) model on carotid artery segmentation, increasing the AUC from 0.945 to 0.985.

Conclusion

Speckle2Self provides a robust solution for ultrasound denoising that bypasses the need for clean training data. By leveraging scale-invariance as a supervisory signal, it effectively disentangles anatomy from deterministic speckle. While the current implementation can introduce slight blurring due to the downsampling operations, future work focusing on raw Radio-Frequency (RF) data and physics-informed priors promises to further refine image sharpness and diagnostic clarity.

Summary updated

BibTeX
@article{li2025speckle,
  author = {Xuesong Li and Nassir Navab and Zhongliang Jiang},
  title = {{Speckle2Self: Self-Supervised Ultrasound Speckle Reduction Without Clean Data}},
  journal = {Medical Image Analysis},
  year = {2025},
  doi = {10.1016/j.media.2025.103755},
}
Figure