Energy-regularized self-supervised deep network for speckle noise reduction in optical coherence tomography

Optical Coherence Tomography (OCT) speckle degrades the contrast and impairs automated analysis. Classical speckle suppression methods blur boundaries, supervised deep learning requires clinically impractical training pairs, and existing self-supervised approaches, such as S2SNet, tend to over-smooth retinal layers. To address these limitations, this study proposes a novel, to the best of our knowledge, energy-based regularization term that penalizes overly smooth, low-energy reconstructions, yielding an Energy-regularized Self-Supervised Network (E-S2SNet). Evaluation of the 18 B-scan SD-OCT dataset and the NR206 dataset demonstrated that E-S2SNet achieves a superior balance between speckle suppression and structural preservation, outperforming both classical and deep learning baselines without requiring clean reference (ground truth/training) data. Furthermore, improved downstream retinal layer segmentation indicates the practical utility of the proposed approach, showing promise for future integration into automated OCT analysis.