E-Thesis 150 views
Irregular Domain Deep Learning / SACHIN BAHADE
Swansea University Author: SACHIN BAHADE
DOI (Published version): 10.23889/SUthesis.65003
Abstract
In recent years, the use of machine learning and deep learning on graph data has increased significantly. Convolutional neural networks have achieved remarkable success with grid-like data such as images, but encounter enormous difficulties when learning from more general structures such as graphs....
| Published: |
Swansea, Wales, UK
2023
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Xie, Xianghua. and Edwards, Mike. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa65003 |
| Abstract: |
In recent years, the use of machine learning and deep learning on graph data has increased significantly. Convolutional neural networks have achieved remarkable success with grid-like data such as images, but encounter enormous difficulties when learning from more general structures such as graphs. The inclusion of trainable local filters enables the automated extraction of high-level features in a regular domain; however, due to the irregular structure of graph data, regular domain deep learning operations are constrained. Thus, the development of deep learning models that can successfully learn from graph data is a promising research field with a great potential for impact. In this work, we are looking to generalise deep learning model for learning spatial feature representations on irregular domain topologies for the purpose of learning high-level model (structures, topology, parts) in medical image analysis problem. This thesis focuses specifically on graph neural networks as the primary machine learning model for effectively tacking graph data challenges. This study investigates the use of irregular domains deep learning methods to enhance the high-level model, with applications in medical image segmentation and detection tasks. We first explore how graph construction affects the behaviour of graph convolutional operations. For this purpose, we use a parametrically specified graph to represent a localised sampling operation on an underlying domain, which we subsequently mine for features while analysing the effect of the graph’s construction on the model’s behaviour. After this, we learn features using advanced deep-learning, spectral, and spatial-based graph signal processing techniques for cell segmentation on immunostained images and present a comparative study. Third, we explore the problem of object recognition in an irregular environment. Conventional convolutional neural networks may process Euclidean data for object detection tasks in a number of ways, but the use of graphs to sample from image data requires special consideration, which has the possibility of generalising to non-Euclidean data. We describe a graph convolution based region proposal method for the detection of non-Euclidean data objects. The extraction of a subgraph as a candidate for the prospective object regions is our primary focus. We discovered improvements when comparing our technique to the region based convolutional neural networks method for the Euclidean domain. The last main chapter focuses on the problem of nuclei detection and finds that graph convolutional networks-based cascaded architecture outperforms convolutional neural networks-based techniques and is more stable. In conclusion, we demonstrate the stability of our irregular domain deep learning methods for graph construction, cell segmentation, and nucleus detection applications, as well as its improved performance in comparison to convolutional neural networks-based approaches. |
|---|---|
| Keywords: |
Graph Neural Network, Biomedical Imaging, Machine Learning, Deep Learning |
| College: |
Faculty of Science and Engineering |
| Funders: |
National Overseas Scholarship Scheme for Scheduled Castes, Ministry of Social Justice and Empowerment, India. |