Date of Award

5-2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Biomedical and Chemical Engineering and Sciences

First Advisor

Mehmet Kaya

Second Advisor

James Brenner

Third Advisor

Shaohua Xu

Fourth Advisor

Munevver Mine Subasi

Abstract

With the progressive development of artificial intelligence (AI), automated analysis has been explored for substituting tedious manual methods in cellular image analysis. Additionally, medical information not included in the conventional statistical methods is now being reconsidered with AI-based personalized medicine. The adaptivity and flexibility of AI ensure its efficiency and accuracy. In this work, computer vision and machine learning (ML) methods under the scope of AI are explicitly investigated and applied to cellular images and clinical data for cytological analysis and to provide diagnosis-based assistance/predictions. Firstly, novel computer vision algorithms achieve cell detection and differentiation estimation on phase-contact microscopic images (human cervical cells) to promote automatic cytological analysis. The proposed segmentation algorithm performs better than existing methods (WTS, EGT, and the Ilastik software) in terms of accuracy, computational time, level of automation, flexibility, and practicality. Under the phase-contrast microscope, the undifferentiated cell clusters can be distinguished effectively by the differences in the cell morphology; this susceptibility of differentiation is highly related to tumor genetic heterogeneity. Estimating the differentiation status of cervical cancer cells is critical for understanding cervical carcinogenesis. Normalized image histogram features and the Haralick texture features were used to detect cell morphology and differentiation. Regarding the cell differentiation estimations, the proposed k-means algorithm detects the major undifferentiated cell cluster efficiently confirmed with visual inspection by specialists. Secondly, an ML model (Long Short-Term Memory model) is proposed to predict intracranial pressure for traumatic brain injury patients (CHARIS database) in a short-term period in real time. The Long Short-Term Memory (LSTM) model achieves good accuracy in continuous predictions of the intracranial pressure (ICP) event with a low root mean square error (RMSE) (2.18 mmHg). Finally, data from patients undergoing cardiac catheterization for suspected coronary artery or valvular heart disease and cardiac surgery (DukeCath database) are analyzed via multiple ML methods (such as Gradient Boosting Classifier, Multi-Layer Perceptron Classifier, Random Forest Classifier, Support Vector Machine Classifier, and Linear discriminant Analysis) to identify high risk cardiovascular diseases patients. Within 90 days after the first catheterizations for patients without a history of cardiovascular disease, the Random Forest Classifier is the best model for predicting multiple cardiac incidents; the Gradient Boosting Classifier achieves the most accurate predictions of the receiving of percutaneous coronary intervention (PCI) and coronary artery bypass graft surgery (CABG) treatments. All the proposed methods demonstrated good performance in data analysis. Additionally, with the proposed ML models, early detections of medical conditions are performed accurately, delivering efficient guidance for treatment management.

Comments

Copyright held by author.

Download

Share

COinS