Texture Analysis based Machine Learning Algorithms For Ultrasound Ovarian Tumour Image Classification within Clinical Practices

Al-karawi, Dhurgham (2019) Texture Analysis based Machine Learning Algorithms For Ultrasound Ovarian Tumour Image Classification within Clinical Practices. Doctoral thesis, University of Buckingham.

Al-Karawi, Dhurgham 1305988 - Sept 2019 Thesis_Final_Version_2019.pdf - Accepted Version

Download (4MB) | Preview


Research investigations reported in this thesis, aim to contribute to the efforts of developing reliable ovarian tumour classification tools using texture features extracted from B-mode ultrasound ovarian tumour scan images. This kind of research is necessitated by the shortage of highly trained sonographers and gynaecologists in order to reduce the heavy pressure on healthcare systems throughout the world. Our ultimate aim is to automate the error-prone process of the laborious manual examination of the ultrasound scan images, and we, therefore, exploit advances in Machine learning and computer vision to develop informative software to be integrated within clinical setup. Our research was guided by an extensive literature review of existing research in this and related fields, building on existing collaborations with medical expertise, and evidence from systems biology research that carcinogenesis results in changing the texture of cysts cellular network. These considerations led to adapting image texture analysis approaches as an adequate source for Machine learning algorithms and software tools. Most existing research works in general biomedical image-based diagnostics are directed towards identifying one or few best performing texture features. Instead, our analysis aimed at extracting a suit of texture-based image features that together contribute to effective ultrasound ovarian tumour image classification. This open-minded strategy unearthed a plethora of texture-based features and in different image domains beyond the spatial domain, which depicts a visual image of the scanned tissue. There is a significant variation in the dimensionality of the texture features, included in our investigations, and although we use different well-known classifiers in evaluating performances, the focus of the comparisons made are not on the choice of classifiers. This thesis includes many contributions; the most significant ones can be summarised as follows: 1. Established that even without pre-processing the scanned images spatial domain is a rich source of 7 microscale texture primitives that can distinguish malignant tumour scans from benign ones with accuracy well above being a case of random chance prediction (70% -83%). The simple majority rule fusion of an odd number of features yield accuracy in the range 83% - 90%. 2. Developed a smart adaptive speckle-noise reduction scheme that applies noise reduction in blocks of the cropped tumour images (not the entire image) only if ii(Skewness, Kurtosis) pair in the block satisfies a criterion determined by training. This adaptive pre-processing is shown to significantly improve the performance of all investigated texture schemes, not only the spatial domain ones. 3. Modified the existing frequency domain texture feature (FFGF), by adaptively pre-processing the cropped tumour image prior to computing its Fourier Spectrum, and using a different binarization scheme to extract the bright elliptical shape at the centre of the FFT spectrum. These modifications improved the accuracy of the original FFGF scheme 85.9% to more than 92%. 4. When attempted to reduce the dependencies between the 3 ellipse parameters of FFGF has shown that even better accuracy (> 95%) can be achieved using a single parameter (the minor axes). These results led to establishing that the FFT-spectrum image is a very rich source of texture information only obscured by its somewhat visually “meaningless” display. We found that all of the features extracted from the FFT-spectrum outperform their spatial domain counterpart, and the fusion of the 7 FFT-spectrum based schemes achieved accuracy of > 97.5%. 5. We further extended the list of texture-based image features beyond the spatial domain and beyond the FFGF schemes by extracting some of the previously defined texture features not only from the FFT-spectrum but also in any image transform domain such as the LBP domain. Again, the texture features in the LBP domain outperformed the spatial domain counterparts, with FFGF from the LBP domain achieving accuracy of 94% which even outperforms the modified FFGF. 6. Finally, the extensive experiments simply opened a Pandora Box of image textures. Instead, of continue other image transform domain, we created two versions of an ML-based software that incorporate 9 spatial domain texture-based features (the original 7 + the Skewness + the Kurtosis) to be used for a prospective test of 100 fresh cases, collected and examined histologically by an IOTA expert gynaecologist at Queen Charlotte and Hammersmith Hospital in London during the period (Oct 2018 – Jan 2019). Version 2 incorporates the smart adaptive speckle noise removal resulted in an accuracy of 94%.

Item Type: Thesis (Doctoral)
Uncontrolled Keywords: Machine Learning Algorithms ; Texture Analysis ; Ultrasound Image ; Ovarian Tumour ; Classification Tools
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Divisions: School of Computing
Depositing User: Nicola Button
Date Deposited: 10 Jan 2022 09:40
Last Modified: 10 Jan 2022 09:40
URI: http://bear.buckingham.ac.uk/id/eprint/513

Actions (login required)

View Item View Item