Aortic valve diagnosis and morphology assessment play pivotal roles in cardiovascular health evaluation. This master’s thesis delves into the evaluation of
computer-assisted methods for aortic valve landmark detection in medical im-
ages of the heart by harnessing machine learning techniques, specifically the
spatial configuration network (SCN) and communicative multi-agent reinforcement learning (C-MARL) methods. The research leverages a dataset comprising
120 images of healthy and 40 images of pathological subjects, acquired with the
computed tomography imaging technique. Addressing the need for robust preprocessing, the study introduces atlas registration-based cropping techniques to the
region of interest in order to enhance landmark detection precision. The obtained
results underscore the efficency of the SCN and C-MARL methods. When comparing landmark detection outcomes, SCN achieved a mean Euclidean distance
between the detected and reference landmark locations of 1.14 ± 0.78 mm for
healthy subjects, while C-MARL demonstrated 1.42 ± 0.82 mm. For pathological
subjects, SCN yielded 3.43 ± 6.00 mm and C-MARL achieved 2.66 ± 3.99 mm.
These findings indicate method-specific strengths across different subject categories. The results hold critical implications for clinical practice, enhancing accu-
racy in aortic valve diagnosis and morphology assessment. Furthermore, the integration of the SCN method into a web-based application showcases the practical
applicability of the research. This innovation empowers healthcare professionals across diverse domains with a user-friendly interface to leverage the power
of landmark localization for medical image analysis, underscoring the tangible
impact of cutting-edge research on real-world medical practices.
|
