Author ORCID Identifier

https://orcid.org/0000-0001-6189-3920

Year of Publication

2022

Degree Name

Master of Science in Biomedical Engineering

Document Type

Master's Thesis

College

Engineering

Department/School/Program

Biomedical Engineering

First Advisor

Dr. Abhijit Patwardhan

Abstract

Almost one in every three people worldwide is infected with Toxoplasma gondii (T. gondii). The biology and growth of the parasite’s bradyzoite form in host tissue cysts are not well understood. T. gondii’s metabolic state influences the morphology of its single mitochondrion, which can be visualized using fluorescence microscopy with specific dyes. Hence, fluorescence microscopy images of cysts purified from infected mouse brains carry biological information about bradyzoites, the poorly understood form of the parasite within them. With the help of fluorescence microscopy techniques, previous studies extracted images of the mitochondrion, nucleus, and the inner membrane complex (IMC) providing information on T. gondii’s cysts paving the way for image processing techniques and machine learning to analyze the bradyzoite form of the parasite. Previously, multivariate logistic regression (MLG) was used to classify shapes of mitochondrion. In the present study, in addition to the previously used MLG model, two other machine learning models, Support Vector Machine (SVM) and K Nearest Neighbors (KNN), were used to explore the possibility of better model selection for mitochondrial classification. A minimal model error was used to optimize the classification model performance. Error in any machine learning model is driven by bias, variance, and noise. Through trial and error, the optimal hyperparameters for each model were selected to minimize error. The dataset used consisted of 1940 labeled mitochondrial objects with 22 features, and consisted of five classes Blob, Tadpole, Donut, Arc, and Other. 50% of the dataset was used for training, and the other 50% was used for testing. The overall models’ accuracy of MLG, SVM, and KNN were 79.1%, 78.9%, and 80.3% respectively. Overall classification performance did not vary, but the F score for some classes like Tadpole and Donut showed improvement when using the two newer models. One of the 22 features used was an application of the Histogram of Oriented Gradients (HOG). The HOG feature was replaced with a novel feature that used linear regression of object boundary segments to extract the HOG for only the object’s boundary. The model that used Boundary HOG showed some improvement over the HOG feature. Finally, a new module including a graphical user interface was developed to process and extract shape and intensity information from TgIMC3 images which facilitate further investigations of the parasite biology.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2022.377

Share

COinS