Graduation Year

2023

Document Type

Thesis

Degree

M.S.C.S.

Degree Name

MS in Computer Science (M.S.C.S.)

Degree Granting Department

Computer Science and Engineering

Major Professor

Dmitry Goldgof, Ph.D.

Committee Member

Lawrence O. Hall, Ph.D.

Committee Member

Peter R. Mouton, Ph.D.

Committee Member

Palak Dave, Ph.D.

Keywords

Annotation Automation, Mean Shift, Microscopy Z-stack, Segment Anything Model Adapter(SAM-Adapter), Unbiased Stereology

Abstract

Unbiased stereology refers to a field of applied mathematics \cite{intro-to-stereology} focused on accurate (model and assumption-free) quantification of three-dimensional (3D) objects, typically based on their appearance in 2D sections (planes) through the objects. In the biological sciences, these techniques are widely used for making unbiased estimates of arbitrary-shaped (stochastic) objects such as stained cells, blood vessels, region volumes, etc., in tissue sections through a region of interest (ROI).

This fundamental methodology is widely used for evaluating structural changes that occur in diseases, aging, and pharmaceutical interventions, thereby ensuring reliable outcomes. In terms of limitations, stereology is tedious, time- and labor-intensive, subjective, and requires specialized professionals to manually count (click) hundreds of objects for each case (animal, human). Thus, the methodology is accurate though prone to inter-rater error and low throughput. For these reasons, the global bioscience community needs novel tools capable of achieving accuracy equal to unbiased stereology but without the time, effort, subjectivity, and training required for manual counting.

Our group has been working on developing automatic approaches using machine learning for manual unbiased stereology. However, these techniques are constrained by their inherent need for labeled data to train their models, which is a substantial rate-limiting step to greater progress in the field. The existing methods for 3D cell segmentation cannot be utilized for labeling the z-stacks because of the huge difference in resolution between the z-axis and the x, and y-axes in the z-stacks. Using 2D methods for labeling on the Maximum Intensity Projection(MIP) of the z-stacks would lead to undercounting due to the possibility of overlapping cells in the z-plane (masking) \cite{palakmimounet}. To address this issue, the present study proposes an innovative solution – a mean-shift clustering algorithm in combination with deep learning to produce labels for large z-stack (disector-stack) datasets from a limited pool of labeled data. While this approach requires expert validation, the final results illustrate its effectiveness, with a reported time efficiency improvement of 57\% compared to traditional manual annotation by experts.

Download

Share

COinS