Cell identification and phenotyping using classical machine learning and deep learning

Abstract

Cell identification and phenotyping using classical machine learning and deep learning Recent image analysis and machine learning technologies have enabled the characterization of biology and human health in multiple dimensions, even at the single cell level. By revealing unique characteristics of any cells of interest, single cell analysis offers significant advantages in various fields of science from cancer or brain research through drug development to personalized and targeted treatments. Usually the first step in most microscopy-based single cell analyses task is the identification of nuclei. Precise localization and identification are fundamental for the accurate assessment of important cell functions. Previously the leading approaches were based on classical image processing methods which usually require prior knowledge to accurately fine-tune the parameters assessed. Recently, the revolutionary approach of deep learning has been incorporated into image analysis, and has fundamentally contributed to achieve ground-breaking results in single cell analysis, including image classification, image registration, object detection and recognition, and its applications such as computer-aided diagnosistics. This thesis consists of two parts: phenotypic analysis (with machine learning) and single cell analysis (with deep learning). In the first part, I have focused on answering major questions in phenotypic data analysis, including how to discover the whole dataset effectively or how to increase the accuracy of classification. In the second part the primary focus is on deep learning based approaches. During my research I used and developed various software tools and created algorithms to support in-depth single cell analysis. All of these are aimed to support a higher level of single cell detection and segmentation. This part covers some major research tasks in which I have participated, thus a full automation of a patch-clamping system previously driven manually and a standalone tool to help the user analyze object occurrences on specific regions of an image are presented here. As a result, we have successfully demonstrated how to apply machine learning methods effectively by implementing flexible software tools appropriate for phenotyping. Thereby, we have successfully accomplished the main aim of our research to improve time-efficiency of biological image-based data analysis for human experts. We have also demonstrated that deep learning, a more advanced machine learning technique, utilized in various approaches is capable of finding cells at high precision. Thereby, it has the potential to support automatizing and replacing human tasks in single cell research.