A Novel Statistical-based Approach for 3-D Surface Detection

Smith, Samuel (2021) A Novel Statistical-based Approach for 3-D Surface Detection. Doctoral thesis, Birmingham City University.

Samuel Smith PhD Thesis published_Final version.pdf - Accepted Version

Download (43MB)


The field of medical image analysis is concerned with the extraction of salient information from complex digital imagery. Developments in image acquisition tools has given rise to a number of 2-D and 3-D digital image domains that are capable of mapping the anatomy and internal structures of a patient in an non-invasive fashion. The data produced by these tools is inherently complex, and a number of image processing techniques are commonly applied to simplify the images in order to extract the information deemed most relevant. Image segmentation is one commonly applied process which partitions a digital image into multiple segments that correspond to various structures within the data. The goal of segmentation is to change the representation of an image into something that is easier to visualise and analyse in complex tasks (i.e. targeted clinical treatment planning). Often these segmentation tasks are performed manually by an expert clinician, however the task of drawing object contours is a time consuming process subject to human biases and interpretation. Automated and semi-automated segmentation is a complex, non-trivial process reliant on a number of pre-processing stages which first extract the spatial structural information contained within the image. For 2-D images, the structural information is contained within edge features and there are a number of edge detection algorithms in the literature which have been extensively appraised. For 3-D images this structural information is contained within the surface features, and while surface detection algorithms exist, their development is immature compared to edge detection and formal evaluation in the literature is largely absent. Furthermore, recent developments in statistical methods for 2-D edge feature extraction have showed promise in resolving 2-D structural information in medical data, however no work has yet explored these approaches in 3-D.

In this thesis two novel methods of statistical surface detection are presented, which contribute to the field by transferring approaches of 2-D statistical edge detection into3-D. The proposed methods optimise the resolving power of the 2-D statistical methods while providing accurate surface detection in the x, y and z dimensions of images. The methods are presented with a range of parametric and non-parametric statistical tests which were extensively analysed using both qualitative and objective methods. In addition, the framework for evaluation is an additional novel contribution in this work, which considers individual aspects of surface detection performance, such as the effects of the statistical properties of the regions within the image, the impact of surface topology, and the response to multiple distinct regions present within the image. A comprehensive dataset of controlled interfaces is developed, and performance of the surface detection algorithms were judged using a novel fast implementation of F-measure analysis against ground truth solutions which is new to this work. The surface detection methods are also analysed on real MRI data, and their performance was qualitatively assessed on the ability to detect brain tumour boundaries and structural pathologies in paediatric patient data. For a comparison against the state of art in surface detection, the methods evaluated in the thesis were compared against two existing baseline approaches, namely the 3-D Canny method, and 3-D Steerable filters.

The results of the evaluations reveal that the proposed 3-D statistical method for surface detection offers improved detection of surfaces on synthetic data with varying interface and topology considerations. Furthermore, the proposed methods improve detection of surfaces when the variability in image intensity high, such as within regions of texture, which is suitable for delineating regions of complex structure in MRI data. Additionally, the statistical methods were able to match the performance of the baseline methods under conditions considered optimal for the baseline approaches, such as the detection of surfaces with a strong intensity differential. Furthermore, the proposed methods of surface detection are shown to be suitable for real 3-D data which is anisotropic in resolution, namely on MRI imagery where the z-spacing within the dataset is often of poor resolution. Provided as a recommendation for further work, the best performing techniques are presented, notably these were the χ2 and Student t-test statistical methods. Characteristically these methods produced strong magnitude surfaces with good connectivity on real and synthetic data, with the χ2 test also achieving a good suppression of image noise. Therefore, illustrating the potential of this novel method of 3-D surface detection for medical image analysis applications.

Item Type: Thesis (Doctoral)
26 May 2021Submitted
8 November 2021Accepted
Uncontrolled Keywords: Surface Detection, 3-D Edge Detection, Statistical Filtering, Performance Measures, Segmentation, Multi-modal MRI, Pilocytic Astrocytoma, Paediatrics
Subjects: CAH11 - computing > CAH11-01 - computing > CAH11-01-01 - computer science
Divisions: Doctoral Research College > Doctoral Theses Collection
Faculty of Computing, Engineering and the Built Environment > School of Computing and Digital Technology
Depositing User: Jaycie Carter
Date Deposited: 11 Aug 2022 12:48
Last Modified: 11 Aug 2022 12:48
URI: https://www.open-access.bcu.ac.uk/id/eprint/13484

Actions (login required)

View Item View Item


In this section...