An Optimized Neural Network Architecture for Auto Characterization of Biological Cells in Digital Inline Holography Micrographs

Digital inline holography (DIH) based microscopy is a proven technique for the characterization of biological cells via their diffraction signatures. Most of the prevalent characterization techniques are based on the handcrafted feature extraction methods. This limits the applicability to certain known cell types only. It needs adjustment for every new cell type, whereby features must be manually determined first, making it very tedious and prone to subjective errors. To overcome these problems, we have investigated various representational learning-based artificial neural network (ANN) architectures to classify cell types, namely, red blood cells (RBC), white blood cells (WBC), cancer cells (HepG2 and MCF7), and artificial microbeads. The performance of these ANNs on various dimensions of cell micrographs as well as across other standard machine learning algorithms have been studied to obtain an optimized model and to validate it. This study shows that the convolutional neural network (CNN) based architecture shows a better classification accuracy of similar to 97% as compared to the traditional support vector machine (SVM) based architecture with an accuracy of similar to 71%. These results are comparable to that of the analytical model, which shows the average classification accuracy of similar to 95%. Further, we can incorporate this trained model in the on-board computer of DIH based lens-free microscope to facilitate a portable telemedicine diagnosis device.