Song, Sifan
(2023)
Advanced Representation Learning for Dense Prediction Tasks in Medical Image Analysis.
PhD thesis, University of Liverpool.
|
Text
201026309_Oct2023.pdf - Author Accepted Manuscript Download (13MB) | Preview |
Abstract
Machine learning is a rapidly growing field of artificial intelligence that allows computers to learn and make predictions using human labels. However, traditional machine learning methods have many drawbacks, such as being time-consuming, inefficient, task-specific biased, and requiring a large amount of domain knowledge. A subfield of machine learning, representation learning, focuses on learning meaningful and useful features or representations from input data. It aims to automatically learn relevant features from raw data, saving time, increasing efficiency and generalization, and reducing reliance on expert knowledge. Recently, deep learning has further accelerated the development of representation learning. It leverages deep architectures to extract complex and abstract representations, resulting in significant outperformance in many areas. In the field of computer vision, deep learning has made remarkable progress, particularly in high-level and real-world computer vision tasks. Since deep learning methods do not require handcrafted features and have the ability to understand complex visual information, they facilitate researchers to design automated systems that make accurate diagnoses and interpretations, especially in the field of medical image analysis. Deep learning has achieved state-of-the-art performance in many medical image analysis tasks, such as medical image regression/classification, generation and segmentation tasks. Compared to regression/classification tasks, medical image generation and segmentation tasks are more complex dense prediction tasks that understand semantic representations and generate pixel-level predictions. This thesis focuses on designing representation learning methods to improve the performance of dense prediction tasks in the field of medical image analysis. With advances in imaging technology, more complex medical images become available for use in this field. In contrast to traditional machine learning algorithms, current deep learning-based representation learning methods provide an end-to-end approach to automatically extract representations without the need for manual feature engineering from the complex data. In the field of medical image analysis, there are three unique challenges requiring the design of advanced representation learning architectures, \ie, limited labeled medical images, overfitting with limited data, and lack of interpretability. To address these challenges, we aim to design robust representation learning architectures for the two main directions of dense prediction tasks, namely medical image generation and segmentation. For medical image generation, the specific topic that we focus on is chromosome straightening. This task involves generating a straightened chromosome image from a curved chromosome input. In addition, the challenges of this task include insufficient training images and corresponding ground truth, as well as the non-rigid nature of chromosomes, leading to distorted details and shapes after straightening. We first propose a study for the chromosome straightening task. We introduce a novel framework using image-to-image translation and demonstrate its efficacy and robustness in generating straightened chromosomes. The framework addresses the challenges of limited training data and outperforms existing studies. We then present a subsequent study to address the limitations of our previous framework, resulting in new state-of-the-art performance and better interpretability and generalization capability. We propose a new robust chromosome straightening framework, named Vit-Patch GAN, which instead learns the motion representation of chromosomes for straightening while retaining more details of shape and banding patterns. For medical image segmentation, we focus on the fovea localization task, which is transferred from localization to small region segmentation. Accurate segmentation of the fovea region is crucial for monitoring and analyzing retinal diseases to prevent irreversible vision loss. This task also requires the incorporation of global features to effectively identify the fovea region and overcome hard cases associated with retinal diseases and non-standard fovea locations. We first propose a novel two-branch architecture, Bilateral-ViT, for fovea localization in retina image segmentation. This vision-transformer-based architecture incorporates global image context and blood vessel structure. It surpasses existing methods and achieves state-of-the-art results on two public datasets. We then propose a subsequent method to further improve the performance of fovea localization. We design a novel dual-stream deep learning architecture called Bilateral-Fuser. In contrast to our previous Bilateral-ViT, Bilateral-Fuser globally incorporates long-range connections from multiple cues, including fundus and vessel distribution. Moreover, with the newly designed Bilateral Token Incorporation module, Bilateral-Fuser learns anatomical-aware tokens, significantly reducing computational costs while achieving new state-of-the-art performance. Our comprehensive experiments also demonstrate that Bilateral-Fuser achieves better accuracy and robustness on both normal and diseased retina images, with excellent generalization capability.
| Item Type: | Thesis (PhD) |
|---|---|
| Uncontrolled Keywords: | Representation Learning, Medical Image Segmentation, Medical Image Generation, Chromosome Straightening, Fovea Localization |
| Divisions: | Faculty of Science and Engineering > School of Electrical Engineering, Electronics and Computer Science |
| Depositing User: | Symplectic Admin |
| Date Deposited: | 15 Jan 2024 16:14 |
| Last Modified: | 15 Jan 2024 16:15 |
| DOI: | 10.17638/03176521 |
| Supervisors: |
|
| URI: | https://livrepository.liverpool.ac.uk/id/eprint/3176521 |
Dimensions
Dimensions
