PhD thesis defense to be held on December 17, 2025, at 12:00 (Tele-education Room - Library Building)
Thesis title: ADVANCED METHODS FOR THE ANALYSIS AND PROCESSING OF CARDIOLOGICAL DATA
Abstract: The demand for timely prevention and diagnosis of key human cardiovascular diseases has led to the need for developing advanced techniques for analyzing and processing cardiological data, in order to predict heart diseases before serious symptoms emerge, as well as enabling early diagnosis to improve survival rates and reduce the need for surgeries or invasive procedures. Equally important in advancing these analytical techniques has been the need for increased accuracy in diagnosing and analyzing a patient's cardiological profile, minimizing the margin of error and identifying details that may not be visible through more conventional methods. This PhD thesis focuses on the cardiological data processing, specifically the analysis of high-resolution images derived from Computed Tomography (CT) scans, with the goal of extracting valuable information about the structure and morphology of the heart. This study involves the development and application of specialized algorithms that enhance the accuracy and effectiveness of image segmentation, as well as the data extracted from them. Initially, an introduction of the heart anatomy and imaging methods is provided. Studying the anatomy and imaging techniques of the heart is a fundamental area in cardiology and medical imaging, as it offers crucial insights into the heart's structure, function, and health. Simultaneously, the Imaging and Diagnostic Evaluation Methods of the heart are presented and analyzed, as they are vital for understanding its structure, function, and health. Following that, in this PhD thesis, we present an image pre-processing and segmentation system for the aorta which is a critical process in cardiology that allows precise analysis and evaluation of the aorta to identify pathological conditions, such as aneurysms and atherosclerotic plaques. By combining machine learning and image processing techniques, this system enhances the accuracy and efficiency of diagnostic procedures. Afterwards, the process of segmenting the Aorta in 3D CT data is presented, combining Image Processing and Machine Learning Techniques. This method is highly useful in clinical practice for diagnosing various pathologies, such as aortic dissection and aneurysms. Specifically, a new fully automated 3D segmentation method is implemented on patient data. Initially a thresholding process with a fixed upper and lower threshold T across the image is being used, followed by a classification approach based on a Markov Random Field (MRF) network. The proposed methodology achieved superior segmentation performance compared to all other classical segmentation techniques regarding the accuracy of the extracted 3D aorta model. As a result, the proposed segmentation method can be applied in clinical practice, such as in treatment planning and evaluation, as it can accelerate the assessment of medical imaging data, which is usually a time-consuming and labor-intensive process. Thereafter, this doctoral dissertation provides a detailed Comparative Study of Deep Learning Architectures for the Automatic Segmentation of the Abdominal Aorta in 3D CT data. This study evaluates qualitatively and quantitatively the 3D models results obtained from automatic segmentation using data from both healthy and patient subjects. Comparative results showed that the U-Net deep learning architecture achieved the highest segmentation accuracy among the other three architectures tested. Moreover, U-Net's ability to balance computational efficiency with segmentation accuracy makes it a valuable tool in clinical settings, where timely and precise image analysis is crucial. Finally, in the present doctoral dissertation, we provide an improved segmentation model through the aforementioned Deep Learning Architectures, leveraging the property of pre-training and employing the Intensity Guided Masking (IGM) method, in which regions of the CT image within specific intensity ranges are masked so that the enhanced model is tasked with predicting or reconstructing the covered region. The proposed enhanced method was evaluated both qualitatively and quantitatively, surpassing both advanced supervised benchmark methods and other approaches based on pre-training.
Supervisor: Professor George Matsopoulos
PhD Student: Christos Mavridis