Graduation Year

2023

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Computer Science and Engineering

Major Professor

Dmitry Goldgof, Ph.D.

Co-Major Professor

Lawrence O. Hall, Ph.D.

Committee Member

Peter R. Mouton, Ph.D.

Committee Member

Sudeep Sarkar, Ph.D.

Committee Member

Ashwin Parthasarathy, Ph.D.

Keywords

Unbiased Stereology, Convolutional Neural Networks, Image Classification, Image Detection, MIMO YOLO, Microscopy

Abstract

Quantification of the true number of stained cells in specific brain regions is an important metric in many fields of biomedical research involving cell degeneration, cytotoxicology, cellular inflammation, and drug development for a wide range of neurological disorders and mental illnesses. Unbiased stereology is the current state-of-the-art method for collecting the cell count data from tissue sections. These studies require trained experts to manually focus through a z-stack of microscopy images and count (click) on a hundred or more cells per case, making this approach time consuming (~1 hour per case) and prone to human error (i.e., inter-rater variability). Thus, there is strong interest in the development of new computer-aided methods for automatic counting of stained cells in tissue sections. To address this issue, our group and others have developed both hand-crafted and deep learning approaches for segmenting and counting of brain cells at high magnification (100x objective). However, these approaches continue to suffer from two main limitations.

First, current segmentation-based methods for cell counting require substantial time and effort for generating the ground truth from either 2D images or 3D z-axis stacks of microscopy images. As an efficient alternative to segmentation-based automatic cell counting, this dissertation provides a novel deep learning method for cell detection using an entire z-stack of microscopy images as input and outputs bounding boxes containing the automatically detected cells of interest. Compared to the previous segmentation methods, the proposed method avoids the need for ground truth segmentation masks for training the deep learning models, resulting in substantial time savings for producing automatic cell counts with comparable accuracy to current segmentation-based methods.

The second drawback with much of the previous work for obtaining automatic counts using either cell segmentation or cell detection is the time needed for collecting high magnification images of stained microscopic cells. To address this problem, a second innovation in this dissertation work is a deep learning approach that uses image classification for linking the appearance of cells in low-magnification (20x objective) images with their total cell number in the same region of interest as quantified by unbiased stereology at high magnification. This unbiased method automatically classifies cell number into bins and sub-types cases based on different classes, e.g., treatment. Data collection requires only low-magnification images to generate results in less than 1 minute per case, leading to significant reductions in image capture requirements and data collection times.

To conclude, this dissertation details two novel advances, object detection and classification approaches, for stereology-based quantification of stained microstructures in tissue sections. These innovations overcome the two major drawbacks of automatic cell counting methods that limit progress in many fields of basic research and drug development. By reducing the human time and effort for making accurate, reproducible, and efficient cell counts in specific brain regions, these novel approaches have the potential to accelerate progress across a wide range of neuroscience disciplines.

Download

Share

COinS