Segmentation of three-dimensional abdominal CT images with deeplearning methods

Medical imaging techniques, which enable the rapid and accurate evaluation of internal anatomical structures in clinical processes such as diagnosis, treatment planning, and patient follow-up, are an indispensable component of modern medicine. With the advancement of three-dimensional (3D) imaging technologies, segmentation methods that automatically extract meaningful information from these images have gained increasing importance in clinical applications. This thesis aims to perform automatic multi-organ segmentation on 3D abdominal Computed Tomography (CT) images using deep learning-based modern segmentation architectures, accompanied by a comprehensive analysis. Within the scope of the study, contemporary deep learning architectures such as 3D U-Net, UNETR, and Swin-UNETR are compared in detail; the effects of image preprocessing steps, data augmentation techniques, different loss functions, and post-training operations on segmentation accuracy are evaluated. The main focus of the study is on small and structurally complex organs that tend to exhibit lower segmentation performance. Accordingly, advanced methods such as context removal and dataset-specific transfer learning have been employed. This thesis not only compares different model architectures but also presents a holistic approach to reducing clinically significant segmentation errors. The findings are intended to provide a strong foundation for the integration of 3D segmentation systems into clinical practice and to make a qualified contribution to the literature in this field. As a result of this study, the segmentation performances of different architectures have been analyzed in detail on an organ-specific basis. The importance of preserving anatomical context has been emphasized, especially for small and low-contrast organs, and it has been observed that strategies supported by transfer learning yield more consistent results in these challenging cases. Additionally, beyond segmentation performance alone, comprehensive evaluations have been conducted by considering parameters such as application reliability, training stability, and generalizability, highlighting the practical potential of the proposed approaches in real clinical systems. In this regard, the thesis not only provides academic value but also serves as a guiding framework for developing effective and applicable solutions in the field of medical imaging.