Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.731367
Title: Appearance modelling, pathology classification and evidence pinpointing for medical image analysis
Author: Zhang, Qiang
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 2017
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
We propose several methods to address the tasks of appearance representation, variation modelling, landmark detection, pathology classification and evidence pinpointing in medical image analysis. Object class representation is one of the key steps in various medical image understanding techniques. We propose a part-based parametric appearance model built on Gaussian pyramids we refer to as a Deformable Appearance Model (DAP). A DAP models the variability within a population with local translations of multiscale parts and linear appearance variations of the assembly of the parts. The fitting process uses a two-step iterative strategy: local landmark searching followed by shape regularisation. We present a simultaneous local feature searching and appearance fitting algorithm based on the weighted Lucas-Kanade (LK) method. A shape regulariser is derived to calculate the maximum likelihood shape with respect to the prior and multiple landmark candidates from multi-scale parts, with a compact closed-form solution. We apply the DAP for the tasks of variation modelling and landmark detection. To reduce the redundancy in the representation, we further propose to replace the Gaussian pyramids with wavelet pyramids in the DAPs. The new appearance model is referred to as a Wavelet Appearance Pyramid (WAP). Logarithmic wavelets are adopted to decompose the images into pyramidal complementary channels, each of which represents the image with simple textures at a given scale. The complementary property of the wavelets allows the reconstruction of the object appearance from the image channels. The Supervised Descent Method (SDM) is adopted to model implicitly the prior knowledge and fit the model to new instances. We apply the WAPs for the tasks of landmark detection and pathology classification. To learn on large scale datasets annotated with only class labels and no landmarks, we propose a weakly-supervised method utilising the theories of sparse learning and stochastic optimisation. We pay attention to identifying which specific regions and features of images contribute to a certain classification. In the medical imaging scenario, these can be the evidence regions where abnormalities are most likely to appear, and the discriminative features of these regions supporting the pathology classification. The learning is weakly-supervised requiring only the pathological labelling of the data by clinicians and no other prior knowledge. It can also be applied to learn the salient description of an anatomy discriminative from background, in order to localise the anatomy before a classification step. We formulate evidence pinpointing as a sparse descriptor learning problem. Because of the large computational complexity, the objective function is composed in a stochastic way and is optimised by the Regularised Dual Averaging (RDA) algorithm. We apply the evidence pinpointing method for the tasks of anatomy localisation and pathology classification. We test our object representation and evidence pinpointing methods on the problem of Lumbar Spinal Stenosis (LSS). We validate the performance of DAPs and WAPs on around 200 studies consisting of routine axial and sagittal MRI scans. Intervertebral sagittal and parasagittal cross-sections are typically inspected for the diagnosis of LSS, we therefore build the appearance models on L3/4, L4/5 and L5/S1 axial cross-sections and parasagittal slices. For the task of landmark detection, experiments validate the performance of the DAPs as promising in terms of convergence range, robustness to local minima and segmentation precision compared with conventional shape and appearance models. A further improvement using WAPs is observed in landmark detection and pathology classification. We validate the evidence pinpointing method on three weakly annotated datasets on 600 axial images. Experiments show that compared with supervised methods trained with labels and landmarks, our method gives favourable results trained on larger scale data with only class labels, which demonstrates the learning ability of our method under weak-supervision.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.731367  DOI: Not available
Keywords: RC Internal medicine ; TA Engineering (General). Civil engineering (General)
Share: