Jiang Ziling, Yu Yajun, Fang Jingde, Zhang Hao, Chu Kaiqin, Smith Zachary J
Department of Precision Machinery and Precision Instrumentation, University of Science and Technology of China, Hefei, Anhui 230026, China.
Medical Artificial Intelligence and Automation Laboratory, University of Texas Southwestern Medical Center, Dallas, Texas 75390, United States.
Anal Chem. 2025 Aug 19;97(32):17380-17388. doi: 10.1021/acs.analchem.5c01367. Epub 2025 Aug 6.
Label-free multidimensional imaging techniques are widely used in biological imaging. Among them, Raman imaging and phase imaging are two representative methods. However, Raman signals are inherently weak, leading to low signal-to-noise ratios (SNR) in rapidly acquired spectral data. Similarly, the imaging speed of phase imaging is constrained by both shot and sensor noise. Postacquisition data denoising, in the form of both "traditional" and deep-learning-based methods, can improve data quality. However, most deep learning-based denoising approaches typically rely on high-SNR data for supervised training and often process each slice of the three-dimensional data separately, which neglects useful correlations along the third dimension. In this study, we propose a denoising method based on the 3D Noise2Void (3D N2V) network, which incorporates all three dimensions during the denoising operation, and does not require extensive, high SNR training data. This method effectively removes noise from Raman hyperspectral and 3D phase imaging data in an unsupervised manner while preserving spectral (λ), axial (), and temporal () correlations. We validate our method on Raman data of yeast cells and phase tomography and dynamic imaging data of COS7 cells. The denoising performance of 3D N2V is compared with other two existing methods, Block Matching and 3D Filtering (BM3D) and 3D Residual Channel Attention Networks (RCAN). Experimental results demonstrate that the 3D N2V network effectively reduces noise while preserving essential information and biological features, improving the limit of detection (LOD), and outperforming existing denoising methods.
无标记多维成像技术在生物成像中被广泛应用。其中，拉曼成像和相成像就是两种具有代表性的方法。然而，拉曼信号本质上很微弱，导致快速采集的光谱数据中信噪比（SNR）较低。同样，相成像的成像速度受到散粒噪声和传感器噪声的限制。采集后的数据去噪，包括“传统”方法和基于深度学习的方法，都可以提高数据质量。然而，大多数基于深度学习的去噪方法通常依赖高信噪比数据进行监督训练，并且常常分别处理三维数据的每一个切片，这忽略了沿第三维的有用相关性。在本研究中，我们提出了一种基于3D Noise2Void（3D N2V）网络的去噪方法，该方法在去噪操作过程中纳入了所有三个维度，并且不需要大量的高信噪比训练数据。该方法以无监督方式有效地从拉曼高光谱和三维相成像数据中去除噪声，同时保留光谱（λ）、轴向（）和时间（）相关性。我们在酵母细胞的拉曼数据以及COS7细胞的相断层扫描和动态成像数据上验证了我们的方法。将3D N2V的去噪性能与其他两种现有方法，即块匹配和三维滤波（BM3D）以及三维残差通道注意力网络（RCAN）进行了比较。实验结果表明，3D N2V网络在保留基本信息和生物学特征的同时有效地降低了噪声，提高了检测限（LOD），并且优于现有的去噪方法。
