A Rapid Identification and Counting Method for Particle Sensing Using Machine Learning

Cell identification and counting is important in all cell-related research steps including but not limited to culture, transformation, transfection, protein expression, etc. for cancer studies, disease diagnostics, bacterial detection, and so on. Although hemocytometry has been a traditional and conventional method of counting cells, it is tedious and laborious in general and often provides inaccurate counting. Flow cytometry is a far advanced and accurate method in cell counting but it requires an expensive and delicate instrument. Additionally, flow cytometry cannot be applicable to in situ cell counting. Image analysis is another method in cell counting, which has recently been developed but often suffers from inaccuracy for nonhomogeneous cell populations. In this work, we propose a rapid cell identification and counting method using image analysis and machine learning. This abstract presents preliminary results of a novel approach for counting spherical particles in a mixed population having spherically- and oddly-shaped particles using a machine learning technique based on convolutional neural networks (CNN). The proposed methodology utilizes segmentation of particles through scanning the input image in order to recognize and count spherically-shaped particles. Convolution layers process input images with respective filters to identify the shapes of the particles. Parameters of the filters (weights) are adjusted by the learning algorithm to extract different features of interest. Subsequently, a pooling procedure is applied for selecting identified features, in which the selected region is represented by the max or the average intensity of the scanned region in the input image. Both convolution and pooling procedures are applied multiple times in order to obtain better recognition and counting. Upon completing this learning process, features are fed into fully connected layers. Threshold segmentation and canny edge detection methods are applied in order to obtain a particle pattern of each selected shape to train the CNN. Particles having two different shapes, i.e., spherically- and oddly-shaped particles, are used for preliminary investigation. Data augmentation is used to create subsets of image data for training. After training the algorithm, an image containing many particles in different shapes was tested in order to assess the performance of learning process. Each particle was detected and segmented by a pixel size of 25 x 25. Preliminary results show that CNN has performed promising performance for identifying spherically shaped particles while massive images are currently being gathered and used for training to improve the accuracy. Additionally, initial findings indicate that machine learning could improve cell recognition and counting in biomedical research including cell population sensing. We intend to include a variety of particle shapes for generic cell identification and counting.