Automatic Cardiac Segmentation of Complex Morphologies, Modalities and Tissues
[eng] Cardiovascular diseases (CVDs) continue to take a significant toll on global health, highlighting the need for more accurate and efficient diagnostic tools. This thesis, titled "Automatic Cardiac Segmentation of Complex Morphologies, Modalities, and Tissues Using Deep Learning," delv...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | Universidad de Barcelona |
| Repositorio: | Dipòsit Digital de la UB |
| OAI Identifier: | oai:diposit.ub.edu:2445/217108 |
| Acceso en línea: | https://hdl.handle.net/2445/217108 http://hdl.handle.net/10803/692852 |
| Access Level: | acceso abierto |
| Palabra clave: | Cardiologia Aprenentatge automàtic Diagnòstic per la imatge Imatges per ressonància magnètica Cardiology Machine learning Diagnostic imaging Magnetic resonance imaging |
| Sumario: | [eng] Cardiovascular diseases (CVDs) continue to take a significant toll on global health, highlighting the need for more accurate and efficient diagnostic tools. This thesis, titled "Automatic Cardiac Segmentation of Complex Morphologies, Modalities, and Tissues Using Deep Learning," delves into complex medical imaging and artificial intelligence (AI) technologies necessary to perform advanced and cutting-edge cardiovascular diagnostics. The groundwork of this work is laid by emphasizing the critical importance of early, precise, and personalized CVD assessment by means of machine learning (ML) and deep learning (DL), in order to evolve from qualitative visual assessments and basic quantitative measures into advanced, quantitative, data- driven insights. The importance of accurate delineation of cardiac structures for a correct assessment of their status and function is crucial to move forward in that direction. The first chapter delves into the right ventricle segmentation within magnetic resonance imaging (MRI) images, highlighting the challenges posed by complex shapes and ill-defined borders. It introduces the M&Ms-2 challenge, a substantial dataset encompassing diverse pathologies, multiple views, and various scanners. The chapter discusses the success of nnU-Net and underscores the value of multi-view approaches, indicating the need for comprehensive cardiac segmentation algorithms. In the second chapter, the focus shifts to late gadolinium enhancement MRI (LGE-MRI) segmentation, crucial for quantifying scar tissue in cardiac patients. The proposed solution leverages generative adversarial networks to create synthetic images, enhancing segmentation accuracy in the presence of scar tissue. Results reveal the potential of multi-sequence model training with synthetic images and data augmentation to outperform traditional methods. The third chapter addresses the segmentation of pathological tissue, specifically scar tissue and edema, within multi-modal cardiac MRI images. The chapter introduces a two-staged approach, involving a stacked BCDU-net for accurate myocardium segmentation and multi-modal pathological region segmentation. Anatomically constrained synthetic data augmentation enriches the model's performances. This thesis represents a pioneering effort to enhance cardiac deep learning-driven segmentation. By tackling the complexities of morphologies, MRI modalities and pathological tissues, this research contributes valuable insights, algorithms, and datasets to such task. |
|---|