This thesis investigates the application of deep learning techniques in image analysis across various domains, focusing on four main themes: feature extraction, classification, segmentation, and... Show moreThis thesis investigates the application of deep learning techniques in image analysis across various domains, focusing on four main themes: feature extraction, classification, segmentation, and integration, demonstrating the transformative potential of these technologies. The research begins by addressing feature extraction and classification challenges in agricultural biotechnology. It introduces efficient deep Convolutional Neural Networks (CNNs) that significantly automate and enhance corn seed classification accuracy, surpassing traditional methods. The second theme uses advanced CNNs to classify the ripeness stages of mulberries, enhancing sorting accuracy and improving post-harvest processing, which potentially increases economic value and eliminates the need for specialist assessments. The third theme applies CNNs for segmenting microscope images, particularly focusing on zebrafish larvae in high-throughput settings, demonstrating their ability to accurately differentiate larvae, which supports high-throughput screening and facilitates biological research advancements. The final theme integrates deep learning with Natural Language Processing (NLP) to refine image captioning techniques, creating more precise, context-aware descriptions beneficial to various fields like biomedical imaging and digital media. Employing state-of-the-art deep learning models, the thesis tackles distinct, challenging problems, setting new benchmarks and paving the way for future research. This work underscores the efficacy of deep learning in enhancing image analysis and processing, revolutionizing multiple industries and fostering significant societal and technological advancements. Show less
The main objective of this thesis is to develop new, accurate and reproducible automated methods for the detection and quantification of lesions in coronary and peripheral X-ray angiograms, which... Show moreThe main objective of this thesis is to develop new, accurate and reproducible automated methods for the detection and quantification of lesions in coronary and peripheral X-ray angiograms, which make it possible to extend the straight segment analysis to analyses of sidebranches and bifurcations. We introduce new methods for the detection of pathlines (Wavepath), the detection of arterial contours (Wavecontour) and the measurement of diameter sizes in straight segments, sidebranches and bifurcations. These methods are designed to increase reproducibility and decrease the influence of user interaction. These new methods are validated extensively in coronary and vascular angiograms, proving their accuracy and reproducibility. Furthermore we developed two new bifurcation models (Y-shape and T-shape) in order to accurately measure the diameters and lesion parameters of an entire bifurcation. The models, including their edge segment analyses, are validated extensively in a clinical validation study in order to assess the inter- and intra-observer variability on pre- and post-intervention data. Overall we can conclude that our goal of improving the QCA analysis and extend it towards the new morphologies and new intervention techniques has been met. Show less
