【摘要】 光学相干层析成像技术凭借无创、高分辨率、三维成像等优势，在临床诊断与生物医学研究领域具有重要的应用价值。为了增强其成像性能，研究者从硬件改进到软件处理两方面开展大量工作，同时引入深度学习算法，旨在克服常规成像面临的散斑噪声、分辨率受限、成像深度不足等问题。本文简要概述了光学相干层析图像处理的发展现状，综述了深度学习算法在提升成像性能方面的最新进展，并对该技术的未来发展方向进行了展望，旨在为医学影像技术研究人员提供参考，助力发展更加智能、可靠的成像技术。更多还原

【Abstract】 Significance Optical coherence tomography(OCT) is widely used in clinical diagnostics and biomedical research due to its noninvasive, high-resolution, and real-time imaging capabilities. However, its imaging performance is constrained by inherent physical limitations, including speckle noise, restricted spatial resolution, and limited penetration depth. Traditional solutions such as hardware upgrades or conventional post-processing algorithms often face challenges of high cost, complex workflows, and limited generality. The integration of deep learning has driven a paradigm shift toward intelligent OCT imaging, aiming to overcome these challenges with data-driven strategies and pave the way for more reliable and efficient diagnostic applications.Progress This review systematically summarizes key advances in which deep learning has significantly improved OCT imaging performance across three critical aspects: speckle denoising, resolution enhancement, and sensitivity improvement.In the field of speckle denoising, the research paradigm has clearly evolved from supervised learning relying on well-prepared clean-noisy image pairs to flexible semi-supervised and self-supervised frameworks. Early studies employed generative adversarial networks and dedicated loss functions to achieve effective noise suppression while preserving critical structural boundaries. The field has been further advanced by methods that reduce or eliminate the dependence on ground-truth data. Techniques such as Noise2 Noise and its variants utilize the statistical properties of noise from repeated scans or within single images, while approaches including Neighbor2 Neighbor and state-of-the-art diffusion models exhibit strong capability for detail-preserving denoising from a single noisy input. A notable trend is the extension to three-dimensional volumetric processing, where richer spatial context enables more robust speckle suppression over diverse clinical datasets covering retinal, skin, and placental villi imaging, thus improving the accuracy of downstream tasks such as segmentation and pathology detection.For resolution enhancement, deep learning goes beyond conventional pixel-level super-resolution. A major research focus is to achieve real optical super-resolution beyond the diffraction limit. In the axial dimension, networks trained on spectrally degraded data effectively recover high-frequency components and enhance resolution beyond the physical limits of the light source. For lateral resolution, especially in defocused regions, deep learning models learn to inverse the blur caused by the limited depth of focus. Promising self-supervised methods work directly on interference fringe signals or embed physical priors such as optical propagation models, enabling resolution recovery without paired training data. These advances are not only perceptual but also reveal clinically important features including fine capillaries, subtle retinal lesions, and clear cellular boundaries in biological tissues, which have been validated in ophthalmology, cardiology, dermatology, and other applications.In terms of sensitivity improvement, deep learning provides a software-based solution to the inherent trade-off between penetration depth and illumination safety. By learning from paired low-exposure and high-exposure datasets, models such as SNR-Net can intelligently enhance the brightness of dark, low-photon regions in deep tissues while suppressing noise, effectively extending the effective imaging depth under safe optical power. This approach is particularly valuable for imaging sensitive structures such as the retina. Furthermore, recent studies combining tissue optical clearing with deep learning present a new route. The contrast improvement enabled by chemical agents offers a strong supervision signal for network training, indicating the future potential of virtual sensitivity enhancement in internal tissues where dye application is impractical.Conclusions and Prospects Deep learning is transforming OCT from a conventional imaging tool into an adaptive and intelligent system that supports real-time performance optimization. Future research should focus on establishing standardized large-scale datasets, developing physics-informed networks for higher efficiency and interpretability, and building comprehensive evaluation metrics tailored to clinical tasks beyond conventional PSNR/SSIM. Achieving real-time processing through lightweight model design is critical for clinical translation in scenarios such as intraoperative navigation. The fusion of deep learning and OCT will give rise to the next generation of intelligent quantitative imaging tools, which not only marks a technological breakthrough but also promises to significantly improve the precision and effectiveness of clinical medicine.更多还原