Quantitative pathology using deep learning

The subfield of pathology known as ‘digital pathology’ encompasses the procedures for image acquisition, organization, distribution, labeling, and computational analysis. Digital pathology is expanding quickly and has already produced a remarkable impact in academia and business1. For research, clinical trials, telehealth, remote image processing, and overall patient treatment, the capacity to quickly transmit and evaluate a considerable volume of pathology data is revolutionary2. The use of artificial intelligence is at the center of this new wave of innovation, dieselizing the astonishing advancements in this increasingly digital climate. This Ph.D. thesis is centered around developing advanced pathology machine learning tools to enhance both the analysis and measurement of pathology samples. This dissertation provides a summary of my main pathology related research results, in chronological order, comprising an evolution in machine learning complexity applied to the quantitative assessment of pathology samples. First, I studied pancreatic ductal adenocarcinoma (PDAC) tissue fiber properties using a segmentation algorithm, which constitutes a good reference point for automating the dissection of specific features in tissue biopsies to derive clinically relevant statistics. After being approached by Abbott Laboratories to help examine the myelin content in certain areas of the brain, we used our research group’s quantitative phase imaging techniques to image 19 piglet brain tissue slides. I applied both manual segmentation schemes and deep learning methods to correlate the tissue properties with related size and diet statistics of the tissue subjects. The results exceeded our expectations in terms of discernment capabilities at the single frame level, a task altogether impossible for an expert histopathologist. Additionally, a series of blood smears slides were imaged and subjected to various deep learning devices to not only bypass the need for the standard Wright’s stain, but automatically delineate and label four major white blood cell groups. Finally, in an effort to simplify the pathology slide scanning procedure altogether, and after various attempts and approaches, we arrived at what is now termed ‘GANscan.’ This is a deep learning microscopy method that enables a thirtyfold increase in whole-slide scanning durations using standard equipment and software. This technique has recently been reviewed and declared a “transformative solution [that] can be used to further accelerate the adoption of digital pathology”3 by the eminent scientist Professor Ozcan in a review article on our technique.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Bioengineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Fanous, Michael John

Contributors dc:contributor

• Anastasio, Mark A
• Insana, Michael
• Sutton, Brad
• Dobrucki, Wawrzyniec

• Computational Imaging
• Microscopy
• Ai

Statement dc:rights

• Copyright 2022 Michael Fanous

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/117805

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F117805