Deep learning techniques in biomedical imaging

Thumbnail Image




Κατάκης, Σοφοκλής

Journal Title

Journal ISSN

Volume Title



Musculoskeletal ultrasound (MSK-US) is a valuable diagnostic tool for the examination of the musculoskeletal system since it enables a detailed visualization of the muscles. Within the last decade, various Computer-Aided Diagnosis (CAD) systems emerged as analytic tools that incorporated a selected set of quantitative features (e.g., first-, second- and higher-order statistical features) to assist the clinical staff in the clinical practice. These systems were usually applied for the detection of clinically important parameters such as muscle echogenicity or for the classification of different pathologies based on extracted radiological features. However, with the recent advancements in Artificial Intelligence, new powerful image processing techniques have emerged with their integration in CAD systems to be at an early level. The main objective of this thesis is the introduction of state-of-the art deep learning techniques to improve the performance of the already existing CAD systems. To achieve this, a wide range of medical applications have been studied in a new, large and diverse database. Specifically, the problems examined in this thesis can be categorized in the following topics: 1. Automated Architectural Analysis of the muscles in MSK-US. 2. Automated Textural Analysis of the muscles in MSK-US. 3. Generation of musculoskeletal ultrasound images. Each of these topics holds its own merit, as they investigate distinct aspects of patients' overall health conditions. However, they are also interconnected, as their collective study can yield more accurate computer diagnostic systems. Automated architectural analysis plays a crucial role in real-time muscle inspection, providing standardized assessments of muscle quality in terms of size, as well as the length and angle of its fascicles. Concurrently, automated textural analysis paves the way for systems that exclusively rely on muscle texture to monitor its health—an important aspect of muscular well-being. Textural patterns have the potential to signal abnormalities or specific muscle conditions, potentially enabling early interventions and preventive measures. It's important to note that the development of such systems requires a significant amount of data, a challenge compounded by stringent privacy and regulatory constraints in the medical field. To address these limitations, the automated generation of realistic musculoskeletal ultrasound images emerges as a practical solution. This approach circumvents these barriers, facilitating the acquisition of substantial data volumes essential for advancing research and system development.



Deep learning, Muskuloskeletal ultrasound, Architectural muscle analysis, Textural muscle analysis, Muskuloskeletal synthetic images