Cell Counting and Detection in Microscopy Images using Deep Neural Network

• Author / Creator

  Yao Xue

• Microscopic image analysis is a broad term that covers the use of digital image processing techniques to process and analyze images obtained from a microscope. It is of significant interest to a number of diverse fields such as medicine, biological research, cancer research, drug testing, etc. A typical example is breast cancer, where the tumor proliferation speed (tumor growth) is an important biomarker indicative of the breast cancer patient’s condition. In practical scenarios, the most common method is to examine histological slides under a microscope based on pathologists' empirical assessments, which can be quite accurate in several cases, but generally is slow and prone to fatigue-induced errors. Among the above research fields, cell detection and cell counting are viewed as the central task or basis of microscopic image analysis. In this thesis, I address automatic cell detection and cell counting using computer vision and machine-learning methods.

  Cell counting is to quantitate the population of specific cells from microscope images. The ability of accurate cell counting is important for precision diagnostics in laboratory medicine. I present a supervised learning framework with Convolutional Neural Network (CNN) and cast the cell counting task as an end-to-end regression framework, where the global cell count is taken as the annotation to supervise training, instead of following an object classification or detection framework. Compared to the idea of counting by detection, the regression framework evades the open and difficult problem of detection or segmentation of individual cells, and is more suitable for the counting task, whose end goal is to acquire the number of object instances. To further decrease the prediction error of counting, I fine-tune several cutting-edge CNN architectures (e.g., Deep Residual Network) into the regression model with Euclidean loss function rather than softmax loss function. As the final output, the proposed approach not only estimates the total number of certain cells in an image but also produces the spatial density prediction, which is able to describe the local cell density of an image sub-region. In many clinical imaging systems, researchers have confirmed that the topographic map that illustrates the cell density distribution is a valuable tool correlated with disease diagnose and treatment. The proposed method is evaluated with several state-of-the-art approaches on three cell image datasets and obtain superior performance.

  As a related task to cell counting, cell detection focuses on localizing a certain type of cells or cellular subunits in microscopy images. Here, not only is the population of target cells of interest, but their locations in microscopy images are also valuable for subsequent biomedical research and clinical diagnosis. I propose a cell detection method based on Convolutional Neural Networks (CNNs) that uses encoding of the output pixel space. For the cell detection problem, the output space is the sparsely labeled pixel locations indicating cell centers. I employ random projections to encode the output space to a compressed vector of fixed dimensions. Then, CNN regresses this compressed vector from the input pixels. Furthermore, it is possible to stably recover sparse cell locations on the output pixel space from the predicted compressed vector using $L_1$-norm optimization. I conducted substantial experiments on several benchmark datasets, where the proposed CNN + CS framework (referred to as CNNCS) achieved the highest—or at least top three—performance in terms of F1-score, compared with other state-of-the-art methods.

  On the basis of the proposed CNNCS model, I further develop an end-to-end trainable model, where the CNNCS model's two key components (a CNN-based regression model and CS-based sparse code predictor) are incorporated into a single network structure. Extensive experiments demonstrate the superior performance of the end-to-end trainable model on several challenging cell detection benchmark datasets.

• Subjects / Keywords

  • Microscopy Images
  • Cell Counting and Detection
  • Deep Neural Network

• Graduation date

  Fall 2018

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/R3CZ32M7B

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.