Deep Learning-based data synthesis with additional attention mechanisms

Read full text Summarise with VIRAT
Authors: Gurbandurdy Dovletov
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This dissertation investigates deep learning-based data synthesis with an emphasis on attention mechanisms tailored for domain-specific applications in medical imaging and particle characterization. In the medical domain, the research focuses on synthesizing pseudo-CT (pCT) images from Magnetic Resonance Imaging (MRI) data to address limitations of conventional Computed Tomography (CT), such as ionizing radiation exposure. Novel network architectures, including Targeted Region Enhancement (TRE), Extra-Supervised Attention Gates (ESGA), and Gradient-weighted Class Activation Mapping Guidance (GCG), are proposed to enhance region-specific accuracy and improve bone structure representation in MRI-to-CT translation tasks. Extensive experimental evaluations demonstrate that these attention-driven networks outperform conventional U-Net and GAN-based approaches in both quantitative metrics and visual fidelity. In the field of particle characterization, the dissertation introduces deep conditional generative models for synthesizing sedimentation boundaries in analytical ultracentrifugation experiments and separation kinetics in analytical centrifugation experiments. By incorporating Physical Constraints (PC) into the generative architectures, the proposed methods achieve physically plausible and controllable data synthesis of sedimentation boundaries based on the input sedimentation coefficient. Experimental evaluations reveal that models equipped with PC outperformed their baseline counterparts (cGAN, cVAE, and cVAE-GAN). Furthermore, conditional generative models were employed to synthesize separation kinetics, enabling significant reductions in experimental cost and duration while supporting downstream tasks such as classification.

Fields of Study & Subjects

Generative Adversarial Networks and Image Synthesis Artificial Intelligence in Healthcare and Education Advanced MRI Techniques and Applications

Publication details

Journal / source
Universitätsbibliographie, Universität Duisburg-Essen
Publication date
18 March 2026
Language
EN
Document type
Thesis