Applications of Deep Learning in Single-Cell Analysis

• Main Author: Hollandi, Réka
• Format: Dissertation
• Language: English
• Published: ProQuest Dissertations & Theses 01-01-2021
• Subjects: Agricultural production; Agriculture; Annotations; Artificial intelligence; Cell culture; Cellular biology; Cytoplasm; Deep learning; Documentation; Labeling; Logic; Machine learning; Microscopy; Neural networks; Semantics; Software
• ISBN: 9798381087482

Biological experiments and medical examinations often include the investigation of biological samples at a cellular level – such as analysing cells or subcellular compartments – in order to draw conclusions and propose a proper diagnosis or treatment. Analysis of single cells and more specifically, their phenotyping can lead to much more accurate description of the studied sample than merely examining the sample as an entity e.g. identifying the tumorous region’s extension surrounded by healthy tissue in a section. Considering the frequency and occurrence of affected cells as well reveals more subtle details about the progression of deviation (e.g. in case of a tumour) or vitality and may provide valuable insight into the functional differences compared to healthy samples. Hence, we focus on samples at the single-cell level.Analysing vast amounts of experimental or patient-related data requires appropriate technology both on the acquisition (automated microscopy) and the processing (i.e. analysis software) part. Both can be aided by high-content solutions, such as high-content microscopes and slide scanners intended for the fast and automatic acquisition of samples, and automated (usually also intelligent) software capable of handling such large data. Such software packages often use deep learning (DL), a technology spread worldwide relatively recently.From the early 2010’s deep learninghas been applied in numerous fields of science including computer vision (e.g. object recognition [1] [2] [3] [4] or autonomous driving [5] [6] [7]), language processing (take speech recognition [8] as example), medicine (such as lesion classification [9]) etc. Despite the first appearance of deep neural networks in the 1960s (e.g. perceptron [10], backpropagation [11] [12] [13], multilayer perceptron [14]), widespread application has been limited for decades due to computational complexity and resource requirements until technological advancements like high-performance GPUs (graphical processing units) were made available commercially bringing about the era of deep learningpowered applications worldwide. The underlying methodology is explained in section 1.1.In this thesis, an assisted annotation tool, AnnotatorJ [15], and a so-called image style transferbased segmentation method, nucleAIzer[16], are proposed. Finally, specific applications are presented [17] [18] for the analysis of single cells using the deep learning-based approaches discussed.1.1. Machine learningAn algorithm’s ability to learn the execution of a task either from a set of examples (supervised) or merely features of a collected dataset provided (unsupervised) is referred to as machine learning (ML). A specific supervised machine learning objective, classification, might be reached via numerous classical algorithms (such as kNN: k nearest neighbour, SVM: support vector machines etc.) using training data as examples with each example comprising of a feature vector corresponding to an object; its analogy in cellular analysis is phenotyping i.e. sorting the cells to distinct phenotypes (e.g. healthy, cancerous, mitotic, apoptotic etc.). Let us briefly discuss how classical machine learning is related to neural networks (NNs) and deep learning (DL, see Figure 1.1.) in the following subsections.