Algebraic, Topological, and Geometric Driven Convolutional Neural Networks for Ultrasound Imaging Cancer Diagnosis

Ghafuri, Jehan Sherwan Ziad (2023) Algebraic, Topological, and Geometric Driven Convolutional Neural Networks for Ultrasound Imaging Cancer Diagnosis. Doctoral thesis, The University Of Buckingham.

[img] Text
Final_Thesis_Jehan.pdf - Submitted Version
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (26MB)


Despite the astonishing successes of Convolutional Neural Networks (CNN) as a powerful deep learning tool for a variety of computer vision tasks, their deployments for ultrasound (US) tumour image analysis within clinical settings is challenging due to the difficulty of interpreting CNN decisions compounded by lack of availability of class labelled “good quality” US tumour image datasets that represent an i.i.d random sample of the unknown population. The use of CNN models pretrained on natural images in transfer learning (TL) mode for US image analysis are perceived to suffer from a lack of robustness to small changes and inability to generalisation to unseen data. This thesis aims to develop a strategy for designing efficient CNN architectures customised for US images that overcome or significantly reduce the above challenges while learning discriminating features resulting in highly accurate diagnostic predictions. We first uncover the significant differences in the statistical contents and spatial distribution of image texture landmarks (e.g. Local Binary Patterns) between US images and natural images. Therefore, we investigate the effects of convolution with random Gaussian filters (RGF) on US image content in terms of spatial and an innovative texture-based entropy, and the spatial distribution of texture landmarks. These effects are determined for US scan images of malignant and benign masses for breast, bladder, and liver tissues. We demonstrate that several pretrained CNN models retrained on US tumour scan images in TL mode achieve high diagnostic accuracy but suffer greatly from a lack of robustness against natural data perturbation and significantly low generalisation rates due to highly ill-conditioned convolutional layer filters. Thus,we investigate the behaviour of the CNN models during the training process in terms of three mathematically linked characterisation of the filters point clouds: (1) the distribution of their condition numbers, (2) their spatial distribution using persistent homology (PH) tools, and (3) their effects on tumour discriminating power of texture landmark PH scheme in convolved images. These results pave the way for a credible strategy to develop high-performing customised CNN architectures that are robust and generalise well to unseen US scans. We further develop a newapproach to ensure equal condition numbers across the different channel wise filters at initialisation, andwe highlight their impact on the PH profiles as point clouds. However, the condition number of filters continues to be unstable during training, therefore we introduce a simple novel matrix surgery procedure depending on singular value decomposition as a spectral regularisation. We illustrate that the PH of different point clouds of RGFs and their inverses are distinct (in terms of their birth/death of connected components and holes in dimensions 0 and 1) depending on variation in their condition number distributions. This behaviour changes as a result of applying SVD-surgery, so that the PH of point cloud of a filter set post SVD-surgery approaches the same shape and connectivity of a point cloud of orthogonal RGFs.

Item Type: Thesis (Doctoral)
Uncontrolled Keywords: Cancer Diagnosis ; Convolutional Neural Networks (CNN) ; ultrasound (US) tumour image analysis ; random Gaussian filters (RGF) ; persistent homology (PH).
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
R Medicine > R Medicine (General)
T Technology > T Technology (General)
Divisions: School of Computing
Depositing User: Freya Tyrrell
Date Deposited: 19 Feb 2024 16:21
Last Modified: 19 Feb 2024 16:21

Actions (login required)

View Item View Item