University of South Florida
Active Deep Learning Method to Automate Unbiased Stereology Cell Counting
Abstract
dc:descriptionCell quantification in histopathology images plays a significant role in understanding and diagnosing diseases such as cancer and Alzheimers. The gold-standard for quantifying cells in tissue sections is the unbiased stereology approach. Unfortunately, in unbiased stereology current practices rely on a well-trained human to manually count hundreds of cells in microscopy images. However, this human-based manual approach is time-consuming, labor-intensive, subject to human errors, recognition bias, fatigue, variable training, poor reproducibility, and inter-observer error. Thus, the lack of high-throughput technology for automating unbiased stereology analyses remains a major obstacle to further progress in a wide range of neuroscience and cancer sub-disciplines. This dissertation provides deep learning methods to automate unbiased stereology cell counting in microscopy images of mice stained brain sections. These methods are based on supervised deep learning, which requires large labeled data sets for training in order to learn to count cells effectively. However, labeling data is time-consuming, labor-intensive, and requires expert knowledge in many fields, such as neuroscience. Therefore obtaining labeled data remains a bottleneck for effective supervised leaning based models. Here, we propose multiple approaches to generate and leverage unlabeled data for training deep learning models. First initial microscopic image labels (cell masks) were created using a handcrafted algorithm called the Adaptive Segmentation Algorithm (ASA), followed by human verification, instead of creating pixel-level cell masks manually. Second, due to the limitations of the generalization ability of ASA and the growing amount of unlabeled data --- due to the advancements and automation of microscopy image acquisition technology such as the Stereologer system, an algorithm to leverage unlabeled data using an active learning approach is proposed. This technique uses a previously trained deep learning model to create masks and query an unlabeled pool of data for masks to be verified by a human and increase the training examples for the deep learning model to improve performance. Moreover, the dissertation presents an experimental study of the effect of the non-reproducibility of deep learning models using two different deep learning libraries. The study shows the various sources of non-reproducibility for deep learning results.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy (Ph.D.)
- Level thesis:degree_level
- Dissertation
- Discipline thesis:degree_discipline
- Engineering Computer Science
- Grantor
- University of South Florida
- Year dc:date
- 2020
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Alahmari, Saeed
- Contributors dc:contributor
-
- Dmitry Goldgof, Ph.D.
- Lawrence Hall, Ph.D.
- Sudeep Sarkar, Ph.D.
- Nasir Ghani, Ph.D.
- Peter R. Mouton, Ph.D.
- Robert Gillies, Ph.D.
Subjects
dc:subject × 9Identifiers
dc:identifier.*- Repository record dc:identifier
- https://digitalcommons.usf.edu/etd/8152