Speeding Up the Line-Scan Raman Imaging of ... - ACS Publications

May 7, 2019 - convolutional neural network (CNN) was then trained by a ..... intensity graphs along the red dotted line of the local zoomed- in areas ...
0 downloads 0 Views 7MB Size
Download PDF
Article Cite This: Anal. Chem. 2019, 91, 7070−7077

pubs.acs.org/ac

Speeding Up the Line-Scan Raman Imaging of Living Cells by Deep Convolutional Neural Network Hao He,† Mengxi Xu,‡ Cheng Zong,‡ Peng Zheng,† Lilan Luo,† Lei Wang,*,† and Bin Ren‡ †

School of Aerospace Engineering, Xiamen University, Xiamen 361005, P. R. China The MOE Key Laboratory of Spectrochemical Analysis and Instrumentation, State Key Laboratory of Physical Chemistry of Solid Surfaces, Collaborative Innovation Center of Chemistry for Energy Materials (iChEM), College of Chemistry and Chemical Engineering, Xiamen University, Xiamen 361005, P. R. China

Downloaded via NOTTINGHAM TRENT UNIV on August 11, 2019 at 06:21:33 (UTC). See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

‡

S Supporting Information *

ABSTRACT: Raman imaging is a promising technique that allows the spatial distribution of different components in the sample to be obtained using the molecular fingerprint information on individual species. However, the imaging speed is the bottleneck for the current Raman imaging methods to monitor the dynamic process of living cells. In this paper, we developed an artificial intelligence assisted fast Raman imaging method over the already fast line scan Raman imaging method. The reduced imaging time is realized by widening the slit and laser beam, and scanning the sample with a large scan step. The imaging quality is improved by a data-driven approach to train a deep convolutional neural network, which statistically learns to transform low-resolution images acquired at a high speed into high-resolution ones that previously were only possible with a low imaging speed. Accompanied with the improvement of the image resolution, the deteriorated spectral resolution as a consequence of a wide slit is also restored, thereby the fidelity of the spectral information is retained. The imaging time can be reduced to within 1 min, which is about five times faster than the state-of-the-art line scan Raman imaging techniques without sacrificing spectral and spatial resolution. We then demonstrated the reliability of the current method using fixed cells. We finally used the method to monitor the dynamic evolution process of living cells. Such an imaging speed opens a door to the label-free observation of cellular events with conventional Raman microscopy.

R

imaging speed using the nonlinear Raman effect,10 which is receiving increasing interest for the label free study of biological samples. Its imaging speed can be further improved by sampling only ∼20% pixels of the three-dimensional spectroscopic image stack and recovering the spectral signatures and concentration maps with a non-negative matrix factorization algorithm.11 However, the spectral window (c.a. 200 cm−1) is yet to be widened, the cost is yet to be reduced, and possible laser damage is yet to be avoided before its wide application. Line-scan Raman microscopy is a promising fast Raman imaging technique receiving increasing interests for imaging single living cells. For instance, Fujita’s group12 used the technique to observe the molecular dynamics in apoptotic cells using the Raman signal of cytochrome c as a signature. Huang et al.13 employed it to study the drug release in living cells. In line scan technique, the sample is illuminated by a line-shaped laser and serval hundreds of spectra from the illuminated line can be acquired simultaneously. Then, the laser line is scanned

aman imaging is a noninvasive and label-free optical method and has played a significant role in understanding many life processes.1−8 Compared with conventional imaging techniques, Raman spectroscopy can obtain molecular fingerprint information from single cells to help a comprehensive understanding of biochemical processes. However, the Raman signals of most biomolecules are so weak that a long acquisition time is required to obtain a clear Raman image, which makes it difficult to track the dynamic biological activities of living samples. Therefore, it is highly desired to speed up the Raman imaging. Wide-field imaging has been proposed to improve the imaging speed. In this way, the sample is illuminated by a large laser spot and the Raman light of a particular frequency from the illuminated area is selected by a narrow band-pass filter and imaged on the charge coupled device (CCD) camera. A widefield Raman imaging system can now achieve a temporal resolution as fast as 20 ms.9 However, this technique has to sacrifice the spectral information on the full Raman spectrum other than the particular frequency of interest, which hinders its application in multicomponent analysis. Alternatively, stimulated Raman scattering (SRS) and coherent anti-Stokes Raman Scattering (CARS) microscopy can achieve a video rate © 2019 American Chemical Society

Received: December 28, 2018 Accepted: May 7, 2019 Published: May 7, 2019 7070

DOI: 10.1021/acs.analchem.8b05962 Anal. Chem. 2019, 91, 7070−7077

Article

Analytical Chemistry

Figure 1. A comparison of the fast speed Raman imaging method to obtain a high-resolution Raman image with the conventional line scan method. (a) Conventional low speed Raman imaging with an optimal set of narrow slit, tight line-focus laser beam, fine scan step, and long exposure time. (b) Fast speed Raman imaging with a wide slit, wide line-focus laser beam, large scan step, and short exposure time. The obtained low-resolution Raman images are input to the CNN, whose output becomes a high-resolution one. Different from the conventional imaging method, the fast speed method is to scan the sample in a way in which the excitation laser beam overlapped with two neighboring lines.

approach using a wide slit (150 μm), expanded laser focus line (1.2 μm, under a 100× magnification objective), a large scan step (0.8 μm), short exposure time, and CNN method to obtain a high-resolution Raman image. Different from the conventional continuous scan mode, we scanned the sample with a different pattern to allow one-third of the laser focus line to overlap with the two neighboring lines along the scanning direction, which can save half of the scanning time. Indeed, each scan line contains three parts of intrinsic information for reconstructing the high-resolution image, which can be extracted with the deep learning method. An advantage of the wide slit is that more Raman light can be captured by the spectrometer, as the signal sensitivity increases linearly with the slit width (for more detailed experimental results see the Supporting Information (SI) Figure S1). This enables us to shorten the exposure time to one-third to obtain equal signal intensity compared with that using the narrow slit. Taking both the scanning time and exposure time into account, our method can achieve six times fast imaging speed of the conventional method, theoretically. Reconstruct High-Resolution Image with CNN. CNN belongs to an artificial neural network, which outperforms other neural networks in dealing with image-related tasks, such as classification,16 reconstruction,17−23 and semantic segmentation.24,25 Thanks to CNN’s power in feature extraction and pattern recognition, use of CNN can significantly enhance the imaging capability of an optical system.26,27 Traditional single image super-resolution CNN aims at reconstructing the subsampled images into high-resolution ones, which is a typically ill-posed problem without an analytical solution. Although the low frequency information on the image can be easily recovered through learning or interpolation, the high frequency information is still hard to retrieve. The reason is that the high frequency information which changed rapidly may be treated as random noise, and is hard to learn. Unlike the subsampling process which discards

over the sample perpendicular to the line to obtain a Raman image containing the hyperspectral information. This technique not only retains the full spectral information on every pixel in the imaging area but also significantly shortens the data acquisition time compared with conventional point-scan Raman microscopy.14 Its imaging speed is yet to be further improved for a wider application in monitoring dynamic biological processes. Pavillon15 reported that the compressed sensing technique can increase the line scan imaging speed by significantly reducing the sampling rate down to 20%. The degraded image spatial resolution was then restored by solving a deconvolution model. However, the sparse sampling discarded a part of the spatial information that would be difficult to retrieve by a deconvolution procedure. Here we present a novel way to speed up the line scan Raman imaging by widening the slit and laser beam, scanning the sample with a large scan step, and employing deep learning. More Raman light can enter the spectrometer through the wider slit to improve the sensitivity, which allows the use of a shorter exposure time to improve the imaging speed. A deep convolutional neural network (CNN) was then trained by a large number of imaging data of living cells to transform the acquired low-resolution images to high-resolution ones. We validated the CNN model by reconstructing high-resolution Raman images on the testing samples as well as a fixed cancer cell. The method was finally used for fast imaging the dynamic evolution of living human cancer cells.

■

METHODS AND MATERIALS Fast Speed Line Scan Imaging. Figure 1 illustrates the schematic comparison between the conventional and our fast Raman imaging methods for obtaining high-resolution Raman images. Figure 1(a) displays the conventional line scan Raman imaging using a narrow slit (50 μm), small scan step (0.4 μm), and long exposure time. We denote it as low speed Raman imaging. However, Figure 1(b) demonstrates our current 7071

DOI: 10.1021/acs.analchem.8b05962 Anal. Chem. 2019, 91, 7070−7077

Article

Analytical Chemistry

Figure 2. Spectral degradation model of using a slit of three times width to scan the sample. In the wide slit mode, the measured intensity of position M at λn is the overlap of three shifted parts of A, B, and C in the narrow slit mode. The degradation model can be simply classified as the equation presented, where m denotes the shifted pixel number, which is related with the slit width (denoted as W), pixel size (denoted as S) of the CCD, and magnification rate (denoted as R) of the spectrometer. The equation can be written as m = W × R/S. In the Raman-11 system, a CzernyTurner spectrometer (500 mm focal length) offers a 1:1 magnification rate and the pixel size of the CCD is 20 × 20 μm2.

Figure 3. Scheme of reconstruction of high-resoution Raman images. (a) The scheme of the CNN model to reconstruct the low-resolution image to a high-resolution one. The CNN model consists of 3 convolution layers. The convolution layers extract feature maps from the low-resolution image successively, and finally outputs a high-resolution one. The output image is expected to be as similar as possible with the true high-resolution one. (b) The architecture of the convolution layer, in which N feature maps are generated by convoluting the input image with N filters. These N feature maps are summed together and output one feature after activation with a Rectified Linear Unit (ReLU). The CNN updates its filters after each iteration until convergent.

part of the information, the low-resolution images produced by our method retain all the spectral information in the sampling

area. This enables us to reconstruct the high-resolution image more effectively. 7072

DOI: 10.1021/acs.analchem.8b05962 Anal. Chem. 2019, 91, 7070−7077

Article

Analytical Chemistry

CNN is a complicated model with a lot of hyper parameters. The tuning of these parameters is of great importance for improving the performance of the model. The depth of the convolution layer, the filter size, and the number of each layer are the foremost factors which should be elaborately designed. A deeper CNN can produce better performance but requires a larger training data set. We found that the filter size of 3 × 3 (f1 = f 2 = f 3 = 3) for each layer could produce a good result. A typical set of filter number for each layer is n1 = 64 and n2 = 128. The initial values of these filters are randomly extracted from a Gaussian distribution with zero mean and standard deviation of 0.001. The initial set up of the biases are 0. Batch size and learning rate are also important hyper parameters for CNN. Batch size is the number of samples that are fed to the CNN for each iteration. The CNN updates its weights and biases after the gradients of all samples in a batch have been calculated. Statistically, the calculated gradients are closer to their global values when there are more samples in a batch. On the contrary, the gradients are noisier at a smaller batch, which will lead to a worse learning efficiency. Although a larger batch size can make the convergence to the global optimal station faster, it may slow down the convergence rate in practice.31,32 Learning rate (η) is used to control the update speed of the weights and biases. For instance, in particular, a weight can be updated according to eq 3, where i is the iteration indices, and L denotes the loss function. It is important to choose a proper learning rate to avoid slow convergence at a small learning rate and divergence at a large learning rate. A typical set of the batch size and learning rate in our experiments are 20 and 0.001, respectively.

A CNN directly learns an end-to-end mapping function between the low- and high- resolution image pair. However, living cells are changing continuously over time and it is unreliable to obtain the corresponding low- and highresolution Raman images through two successive scanning processes. Instead, a low-resolution Raman image generated by a wide slit can be simulated by a summing map of the corresponding high-resolution image generated by a narrow slit. In our experiment, the slit was set to three times of the usual case so that the measured signal intensity at a specific wavelength is the overlap of three neighboring parts with shifted wavenumber in the narrow slit mode. The degradation model can be understood as Figure 2, where m is the shifted pixel number along the wavenumber direction on the CCD, which is about 2 pixels in our case. We then used silicon wafer as the sample to validate our degradation model. The simulation performance can be seen in the SI (please see Figure S2). We generated the low-resolution Raman image according to the degradation model discussed above. Before training, we inserted zeros between every two horizontal lines to let the low-resolution images have the same size as the high-resolution ones. Although the image after zero padding has the same pixel number as the high-resolution one, we still call it a lowresolution image. Figure 3(a) illustrates the schematic diagram of reconstructing the high-resolution image with a CNN, which consists of 3 convolutional layers. The convolution layers extract feature maps from the low-resolution image successively, and finally outputs a high-resolution one. The output image is expected to be as similar as possible to the ground truth of the high-resolution one. Figure 3(b) demonstrates the architecture of the convolutional layers, in which N feature maps are generated by convoluting the input images with N filters. These N feature maps are summed together and only one feature was output after activation with a Rectified Linear Unit (ReLU). ReLU28 can make the convergence much faster and still presents good quality. The output response of the ReLU can be described by eq 1. X denotes the input image matrix, W denotes the filter (or weight), b denotes the bias, and Y denotes the convolutional operation. F(X ) = max(0, W ⊗ X + b)

Δi + 1 = 0.9 × Δi + η ×

r

c

∑∑ i=1 j=1

S(i , j) − S(̂ i , j)

(3)

Data Preparation. All Raman hyperspectral data in this work were acquired by a fast line-scan Raman imaging system (Raman-11, Nanophoton, Figure S3). We collected 295 Raman hyperspectral data sets of living human cancer cells (mainly include: HeLa, CaSki, and TCA-8113A cells) under similar experimental conditions (slit width = 50 μm, scan step = 0.4 μm, exposure time per line = 10 s, laser source = 532 nm, laser power = 50 mW, oil-immersion objective: NA = 1.49, 100× magnification). We then chose the wavenumber region of 2798∼3015 cm −1 (138 data points) to generate corresponding high-resolution and low-resolution Raman image pairs. For the convenience of CNN training, we cropped these Raman images to the same size of 45 × 200 pixels.

(1)

Loss value is a direct indicator of the performance of CNN. It is used to evaluate the inconsistency of the neural network output with the true data, which is usually a non-negative real value function. As image reconstruction is a regression problem, we used mean square error (MSE) (see eq 2) as the loss function. 1 MSE = r×c

∂L , Wi + 1 = Wi + 1 + Δi + 1 ∂Wi

■

RESULTS AND DISCUSSION Results on the Testing Sample. The peak area images were used as the training samples, and the CNN was trained on a NIVIDA TITAN XP GPU. The software was programed based on the Python language and TensorFlow framework (a deep learning software framework developed by Google Inc.). We saved the CNN model that has the minimum loss value on the validation samples, and then the model was used to reconstruct high-resolution Raman images. The loss-iteration curves of the training and validation samples are demonstrated in SI Figure S4. Figure 4 demonstrates the result on the testing sample which has not been used to train the CNN. For comparison, we also presented the result of the linear interpolation. Figure 4(a) is

2

(2)

S denotes the true image matrix, and Ŝ is the output of CNN. CNN learns the prior knowledge from the training samples, and updates its weights (W) and biases (b) to gradually decrease the loss and finally achieves a convergence. The update processes of these network parameters were achieved by the gradient descent algorithm. The gradients of the weights and biases can be calculated through the backpropagation algorithm.29 We used Adam, which is a widely used and highly efficient optimization algorithm in the deep learning area,30 to minimize the loss function to allow fast convergent. 7073

DOI: 10.1021/acs.analchem.8b05962 Anal. Chem. 2019, 91, 7070−7077

Article

Analytical Chemistry

value of the reconstructed and the high-resolution Raman image. The formulae of these indicators can be found in the SI. We evaluate the reconstructed images using the above indicators, and the results are presented in Table 1. We Table 1. Image Quality Evaluation Figure 4(c−b) Figure 4(d−b)

PSNR (dB)

SSIM

MAE

STDAE

33.77 35.36

0.89 0.92

0.0130 0.0116

0.0158 0.0125

observed a similar phenomenon on different testing samples. The CNN can increase ∼5% of PSNR and ∼4% of SSIM, whereas decrease ∼10% of MAE and ∼20% of STDAE compared with the linear interpolation method. More results can be seen in the SI Figure S5. We further showed that the hyperspectral image stack can be recovered in a similar way. In this case, the low-resolution hyperspectral image stacks were split into single wavenumber images and fed into the CNN model, and then the highresolution hyperspectral image stacks were output. We chose the wavenumber window 1408∼1485 cm−1 and 2798∼3015 cm−1 as the target to carry out the following experiments. Figures 5(a−c) are the low-, measured high-, and recovered high- resolution Raman image of a cell at 1443 cm−1, respectively. We further extracted the local mean spectra within the red rectangle area marked on Figure 5(b) and presented them in Figure 5(d). Where the red and blue curve denotes the recovered and measured spectrum, respectively. It can be found that the recovered spectrum perfectly coincided with the measured one, and the full width at half maxima (FWHM) of the two spectra are almost identical. This fact indicates that the recovery is considerably reliable, also the lost spectral resolution as a result of the wide slit can be regained through the CNN. To quantitatively analyze the recovery accuracy, the normalized root-mean-square errors (RMSE) of the recovered Raman images to the measured ones at each single wavenumber were calculated and presented in Figure 5(e). The RMSE fluctuates in an acceptable minor region of 1∼5%. Furthermore, we repeated the same processes with 10 hyperspectral data of different cell samples which have the same size of 55 × 220 pixels, and drew the average RMSE histogram in Figure 5(f). Figure 5(g−m) shows the similar results as those demonstrated in Figure 5(a−f) but in the different wavenumber window of 2798∼3015 cm−1. The average normalized RMSEs are within 4%, which are much better than the 10% RMSE reported in ref 15. Speeding Up the Line-Scan Raman Imaging of Cells. We then performed the conventional low speed and our fast speed line scan imaging of a fixed CaSki cell with Raman-11. The excitation laser source was 532 nm, laser power was 50 mW. The focused laser beam width on the sample was 3 μm under a ×100 oil-immersion objective (NA = 1.49), which can cover a slit width of 300 μm, theoretically. For the sake of keeping the same laser power density, we left the laser beam status unchanged in the experimental setup below. For the low speed high-resolution one, the slit width was set to 50 μm, the exposure time per scan was 10 s, and the scan step was 0.4 μm. Figure 6(a) shows the corresponding high-resolution Raman image of the fixed cell in the spectral range of 2913∼2951 cm−1(C−H region) obtained with an acquisition time of 570 s. For the fast speed imaging, the slit width was set to 150 μm, the exposure time per scan was 4 s, and the scan step was 0.8

Figure 4. Comparison of reconstructing high-resolution Raman images on testing sample with CNN and linear interpolation method. (a) Low-resolution Raman image which is generated according to the degradation model described in Figure 2. (b) High-resolution Raman image. (c) The reconstructed image of (a) through linear interpolation. (d) The reconstructed image of (a) through CNN. The right-down corners of (b−d) are the corresponding local zoomed-in images within the red rectangular areas. The image reconstructed by the CNN shows high similarity with the original image. (e−h) The pixel intensity graphs along the red dotted line of the local zoomed-in areas in image (a−d). As we can see, two nearby features can be differentiated in the high-resolution image (b) and the image reconstructed by CNN (d), while they are merged as one in the low-resolution (a) and linear interpolation image (c).

the low-resolution image which is generated according to the degradation model described in Figure 2. Figure 4(c) is the reconstructed image of 4(a) through linear interpolation. Figure 4(d) is the reconstructed image of 4 (a) through CNN. The insets of Figure 4 (a−d) are the corresponding zoomed-in images within the red rectangular areas. From the zoomed-in images, it can be seen that CNN can recover the detailed image information very well. Figure 4(e−h) are the pixel intensity graphs along the red dotted line of the local zoomedin areas in image (a−d). As we can see, two nearby lipid droplets can be differentiated in the high-resolution image (b) and reconstructed image by CNN (d), whereas they are merged as one droplet in the low-resolution (a) and linear interpolation image (c). To evaluate the similarity of the reconstructed and the original high-resolution image, we used 4 indicators to assess it. These indicators are peak signal-to-noise ratio (PSNR), image structural similarity (SSIM), mean absolute error (MAE), and standard deviation of the absolute error (STDAE). A larger PSNR value indicates a better image quality. SSIM is a positive value between 0 and 1. A larger SSIM value indicates that the target image is more similar to the reference. MAE and STDAE are used to evaluate the absolute error of the pixel intensity 7074

DOI: 10.1021/acs.analchem.8b05962 Anal. Chem. 2019, 91, 7070−7077

Article

Analytical Chemistry

Figure 5. Recovery of hyperspectral Raman image stacks of a cell. Raman image at 1443 cm−1: (a) low-resolution image; (b) original highresolution image; and (c) high-resolution image recovered by CNN. (d) Normalized mean Raman spectra (1408∼1485 cm−1) within the red rectangular area (contains 50 pixels) marked in (b). The blue and red colors indicate the measured and the recovered spectrum, respectively. The recovered spectrum perfectly coincided with the measured one, and the fwhm of the two spectra are almost identical, which implies that the recovered spectrum is considerably reliable. Normalized RMSEs of the recovered to measured Raman images at each single wavenumber in the window of 1408∼1485 cm−1: (e) current sample; (f) average of 10 different cell samples with the same size of 55 × 200 pixels. Raman image at 2951 cm−1: (g) low-resolution image; (h) original high-resolution image; (i) high-resolution image recovered by CNN. (j) Normalized mean Raman spectra (2798∼3015 cm−1) within the black rectangular area (contains 50 pixels) marked in (h). The blue and red colors indicate the measured and the recovered spectrum, respectively. Normalized RMSEs of the recovered to measured Raman images at each single wavenumber in the window of 2798∼3015 cm−1: (k) current sample; (m) average of 10 different cell samples.

Figure 6. Experimental results of fixed (a−c) and living (d−f) CaSki cells (a) and (d) high-resolution Raman images produced by the conventional line scan imaging method. The slit was set to 50 μm, the scan step was 0.4 μm, and the exposure time was 10 s per line. The image size of (a) is 57 × 200 pixels and 37 × 200 pixels for (d). (b) and (e) Low-resolution Raman images produced by fast speed imaging method. The slit was set to 150 μm, the scan step was 0.8 μm, and the exposure time was 4 s per line. The image size is 28 × 200 pixels for (b) and 18 × 200 pixels for (d). (c) and (f) The high-resolution Raman images reconstructed by the CNN with the data of (b) and (e), respectively. The image size is 57 × 200 pixels for (c) and 37 × 200 pixels for (f). In (a) and (b) the cell was fixed by formaldehyde.

μm. Figure 6(b) shows the corresponding low-resolution Raman image of the same cell obtained with an acquisition time of 112 s, which is approximately one-fifth that of Figure 6(a). Then the low-resolution image was input to the CNN to obtain the high-resolution image of Figure 6(c). We compared the reconstructed image, Figure 6(c), with the high-resolution image, Figure 6(a), to check whether the CNN model can

reliably recover the information lost by fast scanning. It can be seen that Figure 6(c) is similar to Figure 6(a) and is much clearer than Figure 6(b), indicating that CNN can be successfully applied to real samples. Then we further carried out the same line scan experiments with a living CaSki cell. The parameters for the low speed and fast speed scanning were the same as those of the previous 7075

DOI: 10.1021/acs.analchem.8b05962 Anal. Chem. 2019, 91, 7070−7077

Article

Analytical Chemistry experiments. Figure 6(d) is the high-resolution image produced by the low speed scanning method, and the total data acquisition time was 370 s. Figure 6(e) is the lowresolution image of the same cell produced by the fast speed scanning method. The total data acquisition time of Figure 6(e) was 72 s, which is one-fifth that of Figure 6(d). Then we used the CNN to transform Figure 6(e) into a high-resolution image shown in Figure 6(f). Surprisingly, we observed a different pattern in Figure 6(f) from that in Figure 6(d). Such a difference can be understood by the dynamic change of intracellular species in time and space due to the activities of life which cannot be observed in the fixed cell. The results show the advantages of the current fast imaging method over that of the low speed imaging method in revealing the dynamic evolution of living cell.

■

AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. ORCID

Lei Wang: 0000-0001-7276-0832 Bin Ren: 0000-0002-9821-5864 Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS This work is funded by the National Natural Science Foundation of China (21633005, 21373173, 21790354, and 21711530704), the Ministry of Science and Technology of China (2016YFA0200601), and the Collaborative Innovation Center of High-End Equipment Manufacturing in Fujian. H.H., M.X., and C.Z. contributed equally to this work.

■

CONCLUSIONS In this work, we demonstrate that a deep neural network can significantly improve the spatial resolution and imaging speed of line-scan Raman microscopy. This data-driven approach is solely based on training a deep convolutional neural network, which statistically learns to transform low-resolution and highspeed input images into highly resolved ones. A larger slit and fewer scan steps were used to improve the imaging speed but resulted in a low-resolution image. A wider slit allows more Raman light to be detected, which reduced the acquisition time for each step. The one-third overlapping of the laser focus line with the two neighboring lines along the scanning direction not only saved half of the scanning time but also laid the foundation for the current method, i.e., the signal from each scan line contained the three parts of intrinsic information for reconstructing the high-resolution image with the CNN method. The CNN model trained with a large amount of Raman imaging data of living cells was successfully used to reconstruct high-resolution Raman images from the lowresolution ones acquired by the fast speed imaging method for fixed and living cells. Benefitting from the elaborate recovery of the hyperspectral Raman image stack, the decayed spectral resolution as a result of wide slit can be restored. The current method exhibits five times faster imaging speed without sacrificing the spectral and spatial resolution. We demonstrated a promising deep learning assisted line-scan Raman imaging method for label-free monitoring of the dynamic processes of living cells and gathering high spatial and temporal molecular information with a subcellular resolution. Moreover, such an artificial intelligence assisted technique can also be applied to a broad range of samples and be borrowed for other spectroscopic techniques. If spectral denoising algorithms33,34 can be further used, then we would expect further improvement in the imaging quality and speed of the current method, which is now ongoing.

■

system; training and validation loss curves; and additional testing results (PDF)

■

REFERENCES

(1) Tu, Q.; Chang, C. Nanomedicine 2012, 8 (5), 545−558. (2) Kong, K.; Kendall, C.; Stone, N.; Notingher, I. Adv. Drug Delivery Rev. 2015, 89, 121−134. (3) Paudel, A.; Raijada, D.; Rantanen, J. Adv. Drug Delivery Rev. 2015, 89, 3−20. (4) Zavaleta, C.; De la Zerda, A.; Liu, Z.; Keren, S.; Cheng, Z.; Schipper, M.; Chen, X.; Dai, H.; Gambhir, S. S. Nano Lett. 2008, 8 (9), 2800−2805. (5) Hartsuiker, L.; Zeijen, N. J. L.; Terstappen, L. W. M. M.; Otto, C. Analyst 2010, 135 (12), 3220−3226. (6) Ando, J.; Palonpon, A. F.; Sodeoka, M.; Fujita, K. Curr. Opin. Chem. Biol. 2016, 33, 16−24. (7) Zheng, X. S.; Hu, P.; Cui, Y.; Zong, C.; Feng, J. M.; Wang, X.; Ren, B. Anal. Chem. 2014, 86 (24), 12250−12257. (8) Mariani, M. M.; Maccoux, L. J.; Matthaus, C.; Diem, M.; Hengstler, J. G.; Deckert, V. Anal. Chem. 2010, 82 (10), 4259−4263. (9) Wang, L.; Dai, Y. Z.; He, H.; Lv, R. Q.; Zong, C.; Ren, B. Dynamic Raman imaging system with high spatial and temporal resolution Rev. Sci. Instrum. 2017, 88 (9), 095110. (10) Ozeki, Y.; Umemura, W.; Otsuka, Y.; Satoh, S.; Hashimoto, H.; Sumimura, K.; Nishizawa, N.; Fukui, K.; Itoh, K. Nat. Photonics 2012, 6 (12), 845−851. (11) Lin, H. N.; Liao, C. S.; Wang, P.; Kong, N.; Cheng, J. X. Light: Sci. Appl. 2017, 7 (5), 17179. (12) Okada, M.; Smith, N. I.; Palonpon, A. F.; Endo, H.; Kawata, S.; Sodeoka, M.; Fujita, K. Proc. Natl. Acad. Sci. U. S. A. 2012, 109 (1), 28−32. (13) Huang, J.; Zong, C.; Shen, H.; Cao, Y. H.; Ren, B.; Zhang, Z. J. Nanoscale 2013, 5 (21), 10591−10598. (14) Palonpon, A. F.; Sodeoka, M.; Fujita, K. Curr. Opin. Chem. Biol. 2013, 17 (4), 708−715. (15) Pavillon, N.; Smith, N. I. Opt. Express 2016, 24 (26), 30038− 30052. (16) Hu, W.; Huang, Y.; Wei, L.; Zhang, F.; Li, H. J. J. o. S. J. Sens. 2015, 2015, 1. (17) Dong, C.; Loy, C. C.; He, K.; Tang, X. In Learning a deep convolutional network for image super-resolution. European Conference on Computer Vision; Springer: 2014; pp 184−199. (18) Dong, C.; Loy, C. C.; He, K. M.; Tang, X. O. Ieee T Pattern Anal. 2016, 38 (2), 295−307. (19) Kim, J.; Lee, J. K.; Lee, K. M. In Proc. Cvpr Ieee. 2016, 1646− 1654. (20) Kim, J.; Lee, J. K.; Lee, K. M. In Proc. Cvpr Ieee. 2016, 1637− 1645.

ASSOCIATED CONTENT

* Supporting Information S

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.analchem.8b05962. The equations of image quality indicators; the relation of Raman signal intensity with the slit width; simulation performance of using narrow slit hyperspectral data to simulate the wide slit data; the schematic of the optical 7076

DOI: 10.1021/acs.analchem.8b05962 Anal. Chem. 2019, 91, 7070−7077

Article

Analytical Chemistry (21) Johnson, J.; Alahi, A.; Fei-Fei, L. In Perceptual losses for realtime style transfer and super-resolution. European Conference on Computer Vision.; Springer: 2016; pp 694−711. (22) Ledig, C.; Theis, L.; Huszar, F.; Caballero, J.; Cunningham, A.; Acosta, A.; Aitken, A.; Tejani, A.; Totz, J.; Wang, Z. H.; Shi, W. Z. In 30th Ieee Conference on Computer Vision and Pattern Recognition (Cvpr 2017). 2017, 105−114. (23) Shi, W. Z.; Caballero, J.; Huszar, F.; Totz, J.; Aitken, A. P.; Bishop, R.; Rueckert, D.; Wang, Z. H. In Proc. Cvpr Ieee. 2016, 1874− 1883. (24) Long, J.; Shelhamer, E.; Darrell, T. Proceedings of the IEEE conference on computer vision and pattern recognition. 2015, 3431− 3440. (25) Girshick, R.; Donahue, J.; Darrell, T.; Malik, J. Proceedings of the IEEE conference on computer vision and pattern recognition. 2014, 580− 587. (26) Göröcs, Z.; Tamamitsu, M.; Bianco, V.; Wolf, P.; Roy, S.; Shindo, K.; Yanny, K.; Wu, Y.; Koydemir, H. C.; Rivenson, Y.et al. A deep learning-enabled portable imaging flow cytometer for costeffective, high-throughput, and label-free analysis of natural water samples Light: Sci. Appl. 2018, 7 (1), 20. (27) Rahmani, B.; Loterie, D.; Konstantinou, G.; Psaltis, D.; Moser, C. Light: Sci. Appl. 2018, 7 (1), 7. (28) Nair, V.; Hinton, G. E. Proceedings of the 27th International Conference on Machine Learning (ICML-10) 2010, 807−814. (29) Dreyfus, S. E. J. Guid Control Dynam. 1990, 13 (5), 926−928. (30) Kingma, D. P.; Ba, J. Adam: A method for stochastic optimization In 3rd International Conference for Learning Representations; San Diego, 2015. (31) Byrd, R. H.; Chin, G. M.; Nocedal, J.; Wu, Y. C. Math Program. 2012, 134 (1), 127−155. (32) Li, M.; Zhang, T.; Chen, Y.; Smola, A. J. In Efficient mini-batch training for stochastic optimization. Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; ACM: 2014; pp 661−670. (33) Huang, C. K.; Hamaguchi, H.; Shigeto, S. Chem. Commun. 2011, 47 (33), 9423−9425. (34) Okuno, M.; Kano, H.; Leproux, P.; Couderc, V.; Day, J. P. R.; Bonn, M.; Hamaguchi, H. Angew. Chem., Int. Ed. 2010, 49 (38), 6773−6777.

7077

DOI: 10.1021/acs.analchem.8b05962 Anal. Chem. 2019, 91, 7070−7077