As an emerging technology, cryo-electron microscopy (cryo-EM) has attracted more and more research interests from both structural biology and computer science, because many challenging computational tasks are involved in the processing of cryo-EM images. An important image processing step is to cluster the 2D cryo-EM images according to their projection angles, then the cluster mean images are used for the subsequent 3D reconstruction. However, cryo-EM images are quite noisy and denoising them is not easy, because the noise is a complicated mixture from samples and hardware. In this study, we design an effective cryo-EM image denoising model, CDAE, i.e., a cascade of denoising autoencoders. The new model comprises stacked blocks of deep neural networks to reduce noise in a progressive manner. Each block contains a convolutional autoencoder, pre-trained by simulated data of different SNRs and fine-tuned by target data set. We assess this new model on three simulated test sets and a real data set. CDAE achieves very competitive PSNR (peak signal-to-noise ratio) in the comparison of the state-of-the-art image denoising methods. Moreover, the denoised images have significantly enhanced clustering results compared to original image features or high-level abstraction features obtained by other deep neural networks. Both quantitative and visualized results demonstrate the good performance of CDAE for the noise reduction in clustering single-particle cryo-EM images.

作为一项新兴技术，冷冻电子显微镜（cryo-EM）吸引了越来越多的结构生物学和计算机科学的研究兴趣，因为冷冻电子显微镜图像的处理涉及许多具有挑战性的计算任务。一个重要的图像处理步骤是根据投影角度对 2D 冷冻电镜图像进行聚类，然后将聚类平均图像用于后续的 3D 重建。然而，冷冻电镜图像噪声很大，去噪并不容易，因为噪声是来自样本和硬件的复杂混合物。在本研究中，我们设计了一种有效的冷冻电镜图像去噪模型，CDAE，即级联去噪自动编码器。新模型由深度神经网络的堆叠块组成，以渐进的方式减少噪声。每个块包含一个卷积自动编码器，通过不同信噪比的模拟数据进行预训练，并通过目标数据集进行微调。我们在三个模拟测试集和一个真实数据集上评估这个新模型。与最先进的图像去噪方法相比，CDAE 实现了非常有竞争力的 PSNR（峰值信噪比）。此外，与原始图像特征或其他深度神经网络获得的高级抽象特征相比，去噪图像具有显着增强的聚类结果。定量和可视化结果都证明了 CDAE 在聚类单粒子冷冻电镜图像中降噪方面具有良好的性能。