-
将空间光场
$ {I^m}(x,y) $ 加载到数字微镜阵列上对空间光场调制。调制后的空间光场经透镜投射至成像样品$ O(x,y) $ ,单像素探测器接收携带物体信息的光场强度信号$ {B^m} $ ,成像原理如图1所示。M次探测后的样品重构图像的二阶关联公式表示如下:$$ G(x,y) = \left\langle {{B^m} \cdot {I^m}(x,y)} \right\rangle $$ (1) 式中:
$ G(x,y) $ 表示光场和探测值的二阶关联函数;$ \left\langle {} \right\rangle $ 代表求系综平均。以矩阵形式表示图像重构过程,将探测值
$ {B^m} $ 转化为一维列向量,即:$$ B = {[ {{B^1}}\;\;\; {{B^2}}\;\;\; \cdots \;{{B^M}} ]^{\rm{T}}} $$ (2) 式中:
$ {\left[ {} \right]^{\rm{T}}} $ 代表矩阵的转置;M代表总探测次数。将物体$ O(x,y) $ 同样转化为一维列向量:$$ O={\left[ \begin{array}{cc}O\left(1\text{,}1\right) \begin{array}{cc}\begin{array}{cc}O\left(1\text{,}2\right)\cdots \end{array} O\left(m\text{,}n\right)\end{array}\end{array} \right]}^{{\rm{T}}} $$ (3) 将用于调制光场的矩阵A表示为:
$$ A=\left[\begin{array}{cccc} I_{1}^{1} & I_{2}^{1} & \cdots & I_{N}^{1} \\ I_{1}^{2} & I_{2}^{2} & \cdots & I_{N}^{2} \\ \vdots & \vdots & \ddots & \vdots \\ I_{1}^{M} & I_{2}^{M} & \cdots & I_{N}^{M} \end{array}\right] $$ (4) 式中:N代表调制矩阵总像素数。所有的探测值向量B即可表示为:
$$ B = AO $$ (5) 由此,公式(1)改写为:
$$ G = \frac{1}{M}{A^{\rm{T}}}B = \frac{1}{M}{A^{\rm{T}}}AO $$ (6) 当m=M,即采样率为100%,重构成像质量较高,但同时伴随着较长的成像时间和较大的存储空间。为了满足实际生物应用,成像系统往往需要通过欠采样的方式缩短成像时间,节约存储空间,但这必将导致重构图像质量下降,图像噪声大幅提升。
-
为了验证系统的多分辨成像能力,测试了不同采样率β下的成像结果:
$$ \beta = \frac{m}{M} $$ (7) 式中:m代表欠采样下的实际测量次数;M代表满采样下的测量次数。以纳米分辨率板(HIGHRES-1,NEWPORT,美国)作为待测目标物体(最小线宽137 nm)。首先,通过100倍物镜(浸油物镜)对分辨率板进行放大,获取了8组不同采样率下的成像结果(图像分辨率:128 pixel×128 pixel),如图4所示。当β=50%时,成像结果质量较差,分辨率版小刻度位置已无法识别。随着采样率增大,成像结果逐渐清晰,图像细节逐渐显现。这种多分辨显微关联成像系统既可以获得用于生物切片观测的高分辨率图像,又能够通过欠采样的方式满足细胞筛选、血流成像等实时成像中低分辨率图像的需要。实际应用中,可以根据对待测目标物体的不同图像质量需求选择不同的分辨率,平衡成像质量与成像时间之间的矛盾关系。
图 4 分辨率板多分辨显微关联成像结果
Figure 4. Imaging results of multi-resolution microscopic correlated imaging for highest resolution target
从图4中可以看出,欠采样方式获取的多分辨成像结果虽然可以大幅缩短成像时间,但其成像质量也显著降低,低采样率的成像结果噪声较大,导致图像细节难以识别,对于生物医学领域的应用带来极大限制。为了解决这一问题,将深度学习与关联成像技术相结合,消除欠采样造成的图像噪声,恢复图像有用信息。
以组织切片(结节性甲状腺肿)作为待测目标物体,测试了经硬件设计和软件设计后的多分辨显微关联成像(DLGI)系统性能。同样获得了5组不同采样率下的成像结果(图像分辨率:128 pixel×128 pixel),如图5所示。随着采样率的降低,系统的成像质量显著下降,同时伴随着大量噪声产生。当采样率达到60%时,传统关联成像(GI)中生物组织的内部细节无法识别,这种图像质量对于病理切片观察是不可接受的。
图 5 不同采样率下生物样本GI和DLGI成像结果对比
Figure 5. Comparison of imaging results of biological samples from GI and DLGI with different sampling rates
将DLGI与GI结果进行对比,获取的重构图像在不同采样率下的重构性能进行比较。采用深度学习方法后,图像质量显著改善。即使在采样率为60%的情况下,生物组织的内部细节和边缘轮廓也会清晰还原,并且图像噪声明显改善。
为了直观地说明成像系统的优势,表1显示了不同采样率下的时间和内存消耗。成像时间主要取决于DMD的调制频率(文中设置DMD调制频率为20 kHz)和深度学习的推理时间。所采用的基于重参数化思想的超高效轻量超分网络显著降低了图像计算复杂度,推理时间可达51 ms。内存消耗包括两部分:预置矩阵的存储空间和探测器探测值的存储空间。每个探测器强度值占8个字节。在传统关联成像方法中,随着采样率的降低,时间和内存消耗虽然逐渐减少,但图像质量大幅下降,无法满足实际需求。然而,所提出的多分辨显微关联成像系统,即使采样率达到60%,也可以识别重构图像的细节。文中的成像系统在保证成像质量的同时,成像时间可节约0.37 s,同时显著减少了内存占用,研究结果对生物医学领域具有重要意义。
表 1 不同采样率下的成像时间和存储空间
Table 1. Time and memory consumption of images with different sampling rates
Sampling rate 100% 90% 80% 70% 60% 50% Sampling time/s 0.745 0.671 0.596 0.522 0.447 0.373 DL time/s 0.051 0.051 0.051 0.051 0.051 0.051 Total imaging time/s 0.796 0.722 0.647 0.573 0.498 0.424 Memory consumption/kB 1 048 704 943 833 838 963 734 092 629 222 524 352 为了验证系统对不同类型样本的适用性,对形态差异较大的生物样本(肝细胞肝癌)进行测试,同样取得了较为理想的成像结果,如图6所示。
为了进一步标定所设计的成像系统的分辨能力,对100倍物镜放大后的分辨率板进行了不同采样率下的成像测试(图像分辨率:128 pixel×128 pixel),如图7所示。
图7(a)为采样率为70%时的GI成像结果,此时分辨率板最小线条位置处(11)已被噪声覆盖,无法识别。但DLGI的图像质量明显提升,噪声消除,如图7(b)所示。提取了GI、DLGI和原始图像目标位置的强度分布(图7(a)~(c)红色虚线,对应图像110像素点位置),并绘制了3种不同条件下的目标位置归一化强度分布曲线。从图7(d)~(f)中可以看出,红色、绿色、蓝色三条虚线标识的位置分别对应分辨率板11-1,11-2,11-3的位置。GI的分布曲线已被噪声覆盖,完全无法识别线条间隔,DLGI的曲线表现出明显的强弱分布规律,与原始图像结果一致,证明了深度学习对系统成像质量的提升能力。
图 6 不同采样率下生物样本GI和DLGI成像结果对比
Figure 6. Comparison of imaging results of biological samples from GI and DLGI with different sampling rates
不同类型生物样本和分辨率板成像结果表明,所设计的多分辨显微关联成像系统具有良好的鲁棒性和抗噪性能,可满足生物医学领域的实时成像需求。
图 7 分辨率板成像结果。(a) GI;(b) DLGI;(c)原始图像; (d)~(f)重构图像特定位置的归一化强度分布图;(g) 表格代表分辨率板不同位置对应的线条间隔(表格来自Newport官网)
Figure 7. Imaging results of resolution target. (a) GI;(b) DLGI;(c) Ground truth; (d)-(f) Normalized intensity distribution of target position in image; (g) The table represents the line spacing corresponding to different positions of resolution target (Newport)
System design of multi-resolution microscopic correlation imaging based on deep learning
-
摘要: 显微成像技术作为研究细胞和生物组织的主要工具,对生物医学的发展起到了极大的推动作用。生物样本的复杂化和生物医学领域对时间和空间分辨率的多样化需求决定了单一功能生物成像系统应用的局限性。为满足生物医学领域的多样化需求,解决成像质量与成像时间之间的矛盾,设计了一种基于深度学习的多分辨显微关联成像系统。该系统通过对显微镜进行硬件设计改造和软件处理,将深度学习与关联成像技术有效结合,当采样率仅为60%时,成像系统能够较好地恢复图像细节,大幅降低欠采样带来的噪声,同时显著提升系统成像的时间分辨率。另外,为了满足所设计的小型多分辨显微关联成像系统的实际需求,采用基于重参数化思想的超高效轻量超分网络,在资源受限的设备下实现实时高质量成像。所提出的成像系统可以在保证成像质量的同时显著缩短成像时间和减少内存占用。不同类型生物样本和分辨率板的测试结果进一步表明了系统的鲁棒性和抗噪性能,研究结果对生物医学领域具有重要意义。Abstract:
Objective Microscopic imaging technology is the primary research method for biological organs, tissues and cells. It plays a significant role in promoting the development of biology and medicine. However, the diversity and complexity of biological samples, the low signal-to-noise ratio, and the optical diffraction limit of traditional optical microscopy significantly limit its application. Different biological samples and different application scenarios have different requirements for microscopic imaging technology. Therefore, in clinical applications, how to obtain images with appropriate resolution through practical needs, and how to shorten imaging time while ensuring imaging quality are the problems that need to be solved urgently in microscopic imaging applications. Methods The microscope is modified by adding a beam-splitting device in the optical path. The light that carries the sample information was exported to the multi-resolution microscopic correlation imaging system after being magnified by the objective lens. The experimental system was integrated into the shell (24 cm×18 cm×12 cm, Fig.2). The optical signal is illuminated to DMD, and the signal light is modulated by DMD and received by a single-pixel detector. The reconstructed images of the sample are obtained through the second-order correlation operation of the modulation matrix and detection intensity of a single-pixel detector. The imaging system was equipped with an industrial computer and a data acquisition card, which are used to control the DMD, load the preset pattern and record the light intensity collected by the detector. The reconstructed images of the sample are obtained through the second-order correlation operation of the modulation matrix and detection intensity of a single-pixel detector. Then the images are processed through deep learning. Results and Discussions The tissue slice was used as the target object, and the performance of the DLGI system after hardware and software design were tested. The imaging results under five different sampling rates were obtained (image resolution: 128×128, Fig.5 and Fig.6). With the decrease of the sampling rate, the imaging quality is reduced significantly, accompanied by a large amount of noise. When the sampling rate reaches 60%, the internal details of biological tissue in traditional correlation imaging (GI) cannot be recognized, and it is unacceptable for pathological section observation. The image quality is significantly improved after using the deep learning method. Even when the sampling rate is 60%, the internal details and edge contours of biological tissues can be restored clearly, and the image noise is significantly improved. In this paper, the ultra-efficient and lightweight hyper-division network based on heavy parameterization reduced the complexity of image calculation significantly (Fig.3), and the reasoning time can reach 51 ms. The imaging time of the imaging system in this paper can save 0.37 s while ensuring the imaging quality and significantly reducing the memory occupation (Tab.1) . Conclusions A multi-resolution microscopic correlated imaging based on deep learning is designed to meet the diverse needs of microscopic imaging and solve the contradiction between imaging quality and imaging time in practical application. The system combines deep learning with correlation imaging technology through hardware design and software processing of the microscope. The imaging system can restore image details with a sampling rate of only 60%, significantly reduce the noise caused by under-sampling, and significantly improve the time resolution of the system. In addition, to meet the actual needs of the small-scale imaging systems, an ultra-efficient super-resolution network is adopted based on the overparameterization method to realize real-time imaging under equipment with limited resources. The proposed imaging system can significantly reduce the imaging time and memory occupation while maintaining imaging quality. The test results of different types of biological samples and resolution boards further show the robustness and anti-noise performance of the system. The research results of the system have great significance for the biomedical field. -
Key words:
- microscopic imaging /
- correlation imaging /
- deep learning /
- multi-resolution imaging
-
图 7 分辨率板成像结果。(a) GI;(b) DLGI;(c)原始图像; (d)~(f)重构图像特定位置的归一化强度分布图;(g) 表格代表分辨率板不同位置对应的线条间隔(表格来自Newport官网)
Figure 7. Imaging results of resolution target. (a) GI;(b) DLGI;(c) Ground truth; (d)-(f) Normalized intensity distribution of target position in image; (g) The table represents the line spacing corresponding to different positions of resolution target (Newport)
表 1 不同采样率下的成像时间和存储空间
Table 1. Time and memory consumption of images with different sampling rates
Sampling rate 100% 90% 80% 70% 60% 50% Sampling time/s 0.745 0.671 0.596 0.522 0.447 0.373 DL time/s 0.051 0.051 0.051 0.051 0.051 0.051 Total imaging time/s 0.796 0.722 0.647 0.573 0.498 0.424 Memory consumption/kB 1 048 704 943 833 838 963 734 092 629 222 524 352 -
[1] Pan Wenhui, Chen Bingling, Zhang Jianguo, et al. Compressed sensing STORM super-resolution image reconstruction based on noise correction-principal component analysis preprocessing algorithm [J]. Chinese Journal of Lasers, 2020, 47(2): 0207024. (in Chinese) doi: 10.3788/CJL202047.0207024 [2] Zhang Saiwen, Deng Yaqi, Wang Chong, et al. Research on super-resolution fluorescence microscopy imaging based on multiple measurement vector compressed sensing [J]. Infrared and Laser Engineering, 2021, 50(11): 20210484. (in Chinese) doi: 10.3788/IRLA20210484 [3] Feng Yueshu, Zhou Cheng, Liu Xuan, et al. Study of multi-resolution microscopic correlated imaging based on Hadamard optimized matrix [J]. Acta Optica Sinica, 2021, 41(21): 116-124. (in Chinese) doi: 10.3788/AOS202141.2111001 [4] Li W W, Tong Z S, Xiao K, et al. Single-frame wide-field nanoscopy based on ghost imaging via sparsity constraints [J]. Optica, 2019, 6(2): 1515-1522. doi: 10.1364/OPTICA.6.001515 [5] Liu H C. Imaging reconstruction comparison of different ghost imaging algorithms [J]. Scientific Reports, 2020, 10(1): 14626. doi: 10.1038/s41598-020-71642-2 [6] Liu Zhentao, Hu Chenyu, Tong Zhishen, et al. Some research progress on the theoretical study of ghost imaging in Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences (Invited) [J]. Infrared and Laser Engineering, 2021, 50(12): 20211059. (in Chinese) doi: 10.3788/IRLA20211059 [7] Lee J, Singh J. Integration of digital twin and deep learning in cyber-physical systems: Towards smart manufacturing [J]. International Journal of Machine Tools and Manufacture, 2020, 38(8): 901-910. doi: 10.1109/TSM.2020.2995548 [8] Kumar J D, Zhang Z, Huang K Q. Multi angle optimal pattern-based deep learning for automatic facial expression recognition-ScienceDirect [J]. Pattern Recognition Letters, 2020, 139: 157-165. doi: 10.1016/j.patrec.2017.06.025 [9] Justine B, Merey R, Silvio G, et al. SeedQuant: A deep learning-based tool for assessing stimulant and inhibitor activity on root parasitic seeds [J]. Plant Physiology, 2021, 186(3): 1632-1644. doi: 10.1093/plphys/kiab173 [10] Zhou Hongqiang, Huang Lingling, Wang Yongtian. Deep learning algorithm and its application in optics [J]. Infrared and Laser Engineering, 2019, 48(12): 1226004. (in Chinese) doi: 10.3788/IRLA201948.1226004 [11] Nehme E, Weiss L E, Michaeli T, et al. Deep-STORM: Super-resolution single-molecule microscopy by deep learning [J]. Optica, 2018, 5(4): 458-464. doi: 10.1364/OPTICA.5.000458 [12] Wang H D, Rivenson Y, Jin Y Y, et al. Deep learning enables cross-modality super-resolution in fluorescence microscopy [J]. Nature Methods, 2018, 16(1): 103-110. doi: 10.1038/s41592-018-0239-0 [13] Lyu M, Wang W, Wang H, et al. Deep-learning-based ghost imaging [J]. Scientific Reports, 2017, 7(1): 17865. doi: 10.1038/s41598-017-18171-7 [14] Wang F, Wang H, Wang H C, et al. Learning from simulation: An end-to-end deep-learning approach for computational ghost imaging [J]. Optics Express, 2019, 27(18): 25560-25572. doi: 10.1364/OE.27.025560 [15] Rizvi S, Cao J, Zhang K, et al. Deringing and denoising in extremely under-sampled fourier single pixel imaging [J]. Optics Express, 2020, 28(5): 7360-7374. doi: 10.1364/OE.385233 [16] Bhardwaj K, Milosavljevic M, Chalfin A, et al. Collapsible linear blocks for super-efficient super resolution[DB/OL]. (2021-03-17) [2022-08-26]. https://arxiv.org/abs/2103.09404v1.