Xu Yuzeng, Wu Guangyi, Yuan Zhuoqun, Li Xinyi, Yang Di, Liang Yanmei
Institute of Modern Optics, Nankai University, Tianjin Key Laboratory of Micro-scale Optical Information Science and Technology, Tianjin 300350, China, Nankai University, Tianjin, 300350, CHINA.
Institute of Modern Optics, Nankai University, Tianjin Key Laboratory of Micro-scale Optical Information Science and Technology, Tianjin 300350, China, Tianjin, Tianjin, 300350, CHINA.
Phys Med Biol. 2025 Jun 25. doi: 10.1088/1361-6560/ade840.
Noise is a key factor determining imaging quality for optical coherence tomography (OCT). Although deep learning has emerged as an effective denoising method, its generalization capability remains limited, especially when test noise levels deviate from the training data.
To solve this problem, we propose a mixed-training strategy combined with a multi-noise level dataset, aiming to enhance model adaptability to unseen noise conditions. The datasets were constructed by inserting different optical attenuators (4 dB, 6 dB, and 10 dB) in the sample arm of an SS-OCT system, simulating diverse noise scenarios. Our mixed-training strategy unified images from 0 dB, 6 dB, and 10 dB into a composite training set, while a portion of the 4 dB data was used as an independent test set to evaluate generalization capability. This approach was applied to classical denoising networks including residual network (ResNet), U-shape network (U-Net), denoising convolutional neural network (DnCNN), and attention-guided deformable convolutional network (ADNet), under both supervised and unsupervised frameworks.
Experimental results demonstrated that models trained with the mixed-training strategy achieved robust performance across noise levels, including unseen 4 dB noise images, where the Unet model trained with the mixed-training strategy under the supervised learning framework attained a PSNR of 29.233 dB and an SSIM of 0.807. This performance is close to that of dedicated models trained on 4 dB data, which achieved a PSNR of 29.221 dB and an SSIM of 0.809. Visual and numerical evaluations further confirmed that the mixed-trained networks effectively suppressed noise artifacts, even under mismatched noise conditions.
This work confirms the value of multi-noise level datasets and introduces a mixed-training strategy that enhances generalization and supports reliable OCT analysis in practice.
噪声是决定光学相干断层扫描（OCT）成像质量的关键因素。尽管深度学习已成为一种有效的去噪方法，但其泛化能力仍然有限，尤其是当测试噪声水平与训练数据不同时。
为了解决这个问题，我们提出了一种结合多噪声水平数据集的混合训练策略，旨在提高模型对未见过的噪声条件的适应性。通过在扫频光学相干断层扫描（SS-OCT）系统的样品臂中插入不同的光衰减器（4 dB、6 dB和10 dB）来构建数据集，模拟不同的噪声场景。我们的混合训练策略将来自0 dB、6 dB和10 dB的图像统一到一个复合训练集中，而4 dB数据的一部分用作独立测试集来评估泛化能力。在有监督和无监督框架下，将这种方法应用于经典去噪网络，包括残差网络（ResNet）、U型网络（U-Net）、去噪卷积神经网络（DnCNN）和注意力引导可变形卷积网络（ADNet）。
实验结果表明，采用混合训练策略训练的模型在不同噪声水平下均表现出稳健的性能，包括未见过的4 dB噪声图像，其中在有监督学习框架下采用混合训练策略训练的U-Net模型达到了29.233 dB的峰值信噪比（PSNR）和0.807的结构相似性指数（SSIM）。这个性能接近在4 dB数据上训练的专用模型，其PSNR为29.221 dB，SSIM为0.809。视觉和数值评估进一步证实，即使在噪声条件不匹配的情况下，混合训练的网络也能有效地抑制噪声伪影。
这项工作证实了多噪声水平数据集的价值，并引入了一种混合训练策略，该策略增强了泛化能力，并在实践中支持可靠的OCT分析。
