Chemical staining in biomedicine is used to observe the phase structure of transparent samples; however, it can change the structure of samples and even kill living cells, making it impossible to observe the life process of samples truthfully under a microscope. Phase contrast overcomes this difficulty and provides a microscopy technique for viewing specimens without labeling or staining, but the costs of complex optical configuration and computational time limit its further application. In addition, image self-focusing is a key step in the process of high-quality phase contrast imaging of samples. Multiple focuses are required to obtain sample information of different regions; thus, cell information cannot be obtained in real time. Current auto-focusing methods based on deep learning are faster and more accurate than conventional auto-focusing methods and do not damage living cells. Therefore, an auto-focusing phase contrast imaging method based on deep learning is proposed in this paper. This can realize auto-focusing phase contrast imaging without manual dyeing, focusing, and other steps. Moreover, a deep learning auto-focus phase contrast imaging (DLFP) network framework is designed for this imaging method, which can realize auto-focus phase contrast imaging quickly and accurately.
Aiming at the problems of manual focusing and phase plate switching in phase contrast imaging of common microscopic images, we proposed an automatic focusing phase contrast imaging method based on deep learning, and designed a network framework based on this method. Considering that with the increase in the depth of the network, the gradient explosion and gradient disappearance occurred during back propagation, we added a residual block (ResNet) on the basis of U-Net to solve the training difficulty caused by network depth and improved the accuracy and precision of the network. Moreover, intensive modules were used to strengthen the reuse of feature information, and the simple, lightweight, Gather-Excite Net spatial attention mechanism was introduced. It can better mine the context information of images or feature map spaces, so that the network can selectively strengthen the useful feature information and suppress useless information, enhance network performance, and improve image resolution. Image acquisition was mainly divided into two parts. The first part was image acquisition, wherein the microscopic image was obtained by manually adjusting the focal length to 0, ±4, ±8, and ±12 μm. This is followed by the obtainment of the corresponding phase contrast image of the microscopic image by moving the circular Zernike phase plate. The second part involved expansion of the dataset by translation, rotation, and other methods. The network input and output image size was 256×256. Approximately 29561 microscopic images at different defocusing distances were used as neural network datasets for different samples in network training, of which 95% and 5% were used as training sets and test sets, respectively.
Experimental test results included random defocused microscopic images of three samples and the corresponding network output results. The high SSIM values of the test output results and real values prove the feasibility of auto-focusing phase contrast imaging method based on deep learning. At the same time, pixel values of DLFP network output images and Ground-truth on horizontal and vertical lines were compared in detail. Three different samples were randomly selected for analysis, and a high degree of coincidence could be achieved in both horizontal and vertical directions, which further proves that the network output image is highly similar to the real one. This also proves the accuracy of the DLFP network. To demonstrate the advantages of the network framework designed in this paper, it was compared with the traditional U-Net structure and the generative adversarial network (pix2pix). DLFP, U-Net, and pix2pix were trained using the same dataset. We found that the pix2pix experimental results contain a lot of noise, whereas those of the U-Net network have lost some details. High SSIM values clearly indicate the advantages of the DLFP network framework. In practical applications, it is difficult to collect a large amount of data, and datasets are very important for deep learning. Therefore, in this paper, the generative adversarial network (pix2pix-GAN), which is often used to complete training with a small number of data sets, is used for experiments to further verify the effectiveness of auto-focusing phase contrast imaging method based on deep learning.
A method based on DLFP network framework to transform defocused micrograph into focused phase contrast image was proposed. The experimental results show that this method can not only effectively improve the identification of cell structural features but also help to understand the phase contrast image in cell analysis and reduce the difficulty and cost of obtaining a phase contrast image. In addition, compared with the U-Net and generative adversarial networks, the high SSIM value shows the advantages of the proposed network framework. Finally, considering the fact that there is too little data in practical application, this paper adopts pix2pix-GAN to verify the validity of auto-focusing phase contrast imaging method based on deep learning, and the experimental results further prove its validity.
