Lightweight Deep Learning Network Assisted Cell Classification Using Lensless Computational Microscopic Imaging Data
Wang Zhaohui,Kang Huan,Chen Duofang,Xu Xinyi,Zeng Qi,Liang Jimin,Chen Xueli
DOI: https://doi.org/10.3788/cjl202249.0507204
2022-01-01
Chinese Journal of Lasers
Abstract:Objective Cell analysis is one of the most important application scenarios of microscopic images, and it plays a vital role in biological research, clinical medicine, drug screening, and other fields. With the rise of artificial intelligence, neural networks have become an important method for cell classification. The neural network has a complex structure, many parameters, a large amount of calculations, and requires advanced graphics cards for training. Taking account of the computer performance of hospitals and biological research institutions, as well as the popularity of mobile terminals today, it is necessary to reduce the computational overhead of neural networks. It can reduce the corresponding medical and scientific research costs, which is conducive to the application and promotion of neural networks for cell classification. It is an urgent need for a method to accurately classify cells in real time and at low cost. Methods Lensless computational microscopic imaging is a low-cost and efficient imaging technology. This imaging method has the characteristics of large field-of-view and high throughput. It can image cells in real time without mechanical focusing or staining. This paper proposes a lightweight network model named Depthwise-ResNeXt to accurately classify cells in real time through the seamless integration of the neural network and lensless computational microscopy. The imaging system (Fig. 1) consists of a laser source, micro-holes and sensors. The laser emitted from the light source passes through the micro-hole and the sample on the holder, and irradiates the sensor. The sensor collects the light signal and transmits it to the computer. The system uses a laser diode as the light source with an output power of 1.5 mW and a wavelength of 532 nm. The diameter of micro-holes is 40 mu m. The signal acquisition uses a CMOS sensor with a pixel size of 1.67 mu m . The distance between the sensor and the sample is within 1 mm. We collect the data of four types of cells including SUM, MCF10A, ECa109, and CL-1 as experimental data. After collected by the sensor, the image data is reconstructed using the angular spectrum method (Equation 1) . The cell picture is extracted by a segmentation algorithm. Specifically, we first reconstruct images at different depths near the focal plane to obtain a pseudo three-dimensional image. Taking advantage of the feature that the edge of the object shows the lowest value in the reconstructed image, we project the lowest point at all depths on the same coordinate to the same plane, so that the cellular object appears black spot-like at the corresponding position. We perform threshold segmentation on the projection image to detect the number and location of the cells, and then cut out the cell images to make a data set. In order to obtain more data, the cell data undergoes data enhancement before entering the network, and the number of data is expanded to 6 times of the original data by means of flipping and rotating. After comprehensive consideration of the amount of calculation, the number of parameters, and the operating speed and accuracy of the experiment, our network chooses ResNet18 as the main body of our network. We try group convolution, channel shuffling, deep convolution, asymmetric convolution kernel, bottleneck structure and other lightweight methods, and conduct the ablation experiments. In the end, our network Depthwise-ResNeXt consists of a convolutional layer, a pooling layer, eight convolutional blocks, a global pooling layer, and three fully connected layers. The numbers of input and output channels of the convolution blocks of the network and the network depth are similar to those of ResNet18. The convolution block borrows the bottleneck structure of ResNeXt. The convolution layer of the main module is replaced by a depthwise convolution. The network learning rate is set to 0.01, and we iterate 300 times. The loss function uses cross entropy loss. The anti-propagation algorithm uses the Adam algorithm. Results and Discussions The classification accuracy of Depthwise-ResNeXt reaches 92.8% . We test the computing capacity of the network with a batch as a benchmark. The computing capacity of the network is 1.01 x 10(9), and the parameter capacity is only 806 kB, which are about 1/10 and 1/14 of those of Resnet, respectively (Table 2) . Compared with other lightweight networks, our network reduces the calculated amount and numbers of parameters to a certain extent with higher accuracy. From the fuzzy matrix of the Depthwise-ResNeXt classification results (Table 3) , it can be seen that the network has the best prediction of CL-1 and the worst recognition of SUM. Conclusions This network proves the possibility of combining neural network and lensless computational microscopy in the field of cell classification, and greatly reduces the application cost of neural networks in cell classification. The cell data in this article only provides simple morphological information. We believe that, with the further improvement of the microscopy technology used, the improvement of data resolution can also bring a qualitative change to the classification accuracy of the network.