Variational and deep learning methods for selective segmentation
Saved in:
| Title: | Variational and deep learning methods for selective segmentation |
|---|---|
| Authors: | Burrows, Liam |
| Summary: | Image segmentation aims to decompose an image into segments in order to determine the position of objects in the image. Naturally this has many applications in a wide range of fields, particularly in medical imaging. This thesis is concerned with the development of reliable selective segmentation methods, tackling difficult challenges such as images containing low contrast or images with inhomogeneous intensity. We begin with developing some models in the variational framework only. We propose a new model which assumes a piecewise-smooth intensity distribution, improving on previous selective segmentation works which usually assume a simple piecewise-constant distribution. This improvement allows our model to effectively segment images displaying intensity inhomogeneity with minimal user input. Our proposed model is a convex selective variant of the famous Mumford-Shah model. Another work focused on the variational framework incorporates reproducible kernel Hilbert space (RKHS) methods to improve edge detection in images containing low contrast. We find that, by modelling edges using approximated Heaviside functions (and smooth parts modelled using kernel functions), that edge detection is improved when compared with using the traditional image gradient information. In the second half of this thesis, we begin to incorporate some deep learning methods. We introduce a joint selective segmentation and registration model and investigate how the regularisation offered by the recent Deep Image Prior (DIP) work can benefit our variational model. As well as this, we dedicate a chapter to combining variational segmentation methods with deep learning methods. Deep learning methods, while incredibly popular in recent years, are often restricted by the requirement of large labelled datasets. We demonstrate how labels can be supplemented (or removed entirely) by using a variational method as a loss function using a semi-supervised (or unsupervised) training algorithm. Overall, we show a range of methods to combine both variational and deep learning segmentation methods. |
| URL: | https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.859213 |
| Database: | OpenDissertations |
| Abstract: | Image segmentation aims to decompose an image into segments in order to determine the position of objects in the image. Naturally this has many applications in a wide range of fields, particularly in medical imaging. This thesis is concerned with the development of reliable selective segmentation methods, tackling difficult challenges such as images containing low contrast or images with inhomogeneous intensity. We begin with developing some models in the variational framework only. We propose a new model which assumes a piecewise-smooth intensity distribution, improving on previous selective segmentation works which usually assume a simple piecewise-constant distribution. This improvement allows our model to effectively segment images displaying intensity inhomogeneity with minimal user input. Our proposed model is a convex selective variant of the famous Mumford-Shah model. Another work focused on the variational framework incorporates reproducible kernel Hilbert space (RKHS) methods to improve edge detection in images containing low contrast. We find that, by modelling edges using approximated Heaviside functions (and smooth parts modelled using kernel functions), that edge detection is improved when compared with using the traditional image gradient information. In the second half of this thesis, we begin to incorporate some deep learning methods. We introduce a joint selective segmentation and registration model and investigate how the regularisation offered by the recent Deep Image Prior (DIP) work can benefit our variational model. As well as this, we dedicate a chapter to combining variational segmentation methods with deep learning methods. Deep learning methods, while incredibly popular in recent years, are often restricted by the requirement of large labelled datasets. We demonstrate how labels can be supplemented (or removed entirely) by using a variational method as a loss function using a semi-supervised (or unsupervised) training algorithm. Overall, we show a range of methods to combine both variational and deep learning segmentation methods. |
|---|
