Appl Opt. 2021 Dec 10;60(35):10964-10974. doi: 10.1364/AO.442678.
Speckle noise is ubiquitous in the optical coherence tomography (OCT) image of the anterior segment, which greatly affects the image quality and destroys the relevant structural information. In order to reduce the influence of speckle noise in OCT images, a denoising algorithm based on a convolutional neural network is proposed in this paper. Unlike traditional algorithms that directly obtain denoised images, the algorithm model proposed in this paper learns the speckle noise distribution through the constructed trainable OCT dataset and indirectly obtains the denoised result image. In order to verify the performance of the model, we compare the denoising results of the algorithm proposed in this paper with several state-of-the-art algorithms from three perspectives: qualitative evaluation from the subjective visual perspective, quantitative evaluation from objective parameter indicators, and running time. The experimental results show that the proposed algorithm has a good denoising effect on different OCT images of the anterior segment and has good generalization ability. Besides, it retains the relevant details and texture information in the image, and it has strong edge preserving ability. The image of OCT speckle removal can be obtained within 0.4 s, which meets the time limit requirement of clinical application.
散斑噪声在眼前节光学相干断层扫描（OCT）图像中普遍存在，这极大地影响了图像质量并破坏了相关的结构信息。为了减少 OCT 图像中的散斑噪声的影响，本文提出了一种基于卷积神经网络的去噪算法。与传统的直接获得去噪图像的算法不同，本文提出的算法模型通过构建可训练的 OCT 数据集来学习散斑噪声分布，并间接获得去噪结果图像。为了验证模型的性能，我们从三个方面将本文提出的算法的去噪结果与几种最先进的算法进行了比较：主观视觉角度的定性评估、客观参数指标的定量评估和运行时间。实验结果表明，该算法对不同眼前节 OCT 图像具有良好的去噪效果，具有良好的泛化能力。此外，它保留了图像中的相关细节和纹理信息，具有很强的边缘保持能力。可以在 0.4s 内获得 OCT 散斑去除图像，满足临床应用的时间限制要求。
