Speckle Denoising in Breast Ultrasound Images Using Multi-Filter Pseudo-Clean Targets and Deep Learning

Ultrasound imaging is widely used in breast cancer diagnosis, but suffers from speckle noise, which reduces contrast and obscures fine structures. Supervised deep learning methods for speckle reduction/denoising typically require clean ground truth, which is unattainable in vivo. To address this, this study proposes a multi-filter pseudo-ground-truth strategy combined with a UNet++ denoiser. Each image in the BUSI dataset is processed using three classical despeckling filters (Gaussian, median, and total variation) to generate diverse pseudo-clean targets. The network is trained with deep supervision to minimize a robust loss with respect to these targets, enabling it to learn a consensus representation beyond any single filter. On the BUSI test set, the proposed method achieves PSNR = 34.11 dB and SSIM = 0.8901, outperforming recent CNN baselines under the same evaluation protocol. Qualitative results show improved edge preservation and lesion visibility. This approach eliminates the need for unattainable clean ultrasound images and provides a practical path toward clinically useful ultrasound despeckling. Code, data splits, pretrained weights, and the full evaluation protocol will be released for reproducibility.

@article{Alani2025,
  title     = {Speckle Denoising in Breast Ultrasound Images Using Multi-Filter Pseudo-Clean Targets and Deep Learning},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {12},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Omar Ayad Alani and Muhammad Moinuddin},
  doi       = {10.14569/IJACSA.2025.0161215},
  url       = {https://doi.org/10.14569/IJACSA.2025.0161215}
}