Deep Learning for Medical Image Analysis: From Single-Modality, Transfer-Modality to Multi-Modality
Guardado en:
| Publicado en: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Autor principal: | |
| Publicado: |
ProQuest Dissertations & Theses
|
| Materias: | |
| Acceso en línea: | Citation/Abstract Full Text - PDF |
| Etiquetas: |
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Resumen: | Medical image analysis plays a critical role in disease diagnosis and treatment planning, offering insights into the structural and functional aspects of the human body. Various imaging modalities—such as magnetic resonance imaging (MRI), computed tomography (CT), and X-ray—serve distinct clinical purposes. For example, Late Gadolinium Enhancement (LGE) MRI is used to detect myocardial scarring, while T2-weighted imaging assesses edema. With recent advances in deep learning, automated and accurate medical image segmentation has become increasingly feasible. My research leverages deep learning to enhance segmentation performance across different imaging modalities and clinical scenarios. I began with single-modality segmentation, working on tasks such as left atrium segmentation using LGE MRI, coronary artery segmentation using X-ray angiography images, and glaucoma screening using color fundus photographs—achieving high accuracy on modality-specific datasets. Building on this foundation, I explored transfer-modality learning through unsupervised domain adaptation (UDA), enabling accurate segmentation in unlabeled target modalities by leveraging knowledge from labeled source domains. More recently, I have focused on multi-modality integration, combining complementary information from multiple imaging modalities while addressing challenges posed by missing or incomplete inputs. To this end, I developed a unified segmentation framework that incorporates a cross-modality feature fusion module and a contrastive learning strategy to better capture intra-subject relationships and fully utilize the available modality information. Looking ahead, I plan to extend my work by developing multimodal and interactive segmentation systems that combine visual data with additional inputs such as textual prompts, audio signals, or point clicks—similar to recent approaches like MedSAM. By incorporating large vision-language models, cross-modal attention mechanisms, and user-in-the-loop interactions, my goal is to build more flexible and adaptive segmentation frameworks. These systems will be able to handle ambiguous cases, reduce the need for manual annotations, and generalize better to different clinical environments. This direction not only aims to improve segmentation performance in complex scenarios but also to close the gap between automated tools and practical clinical use, ultimately supporting more intelligent and interpretable diagnostic systems. |
|---|---|
| ISBN: | 9798291584507 |
| Fuente: | ProQuest Dissertations & Theses Global |