Knowledge Discovery and Machine Learning (KDML)
Knowledge Discovery and Machine Learning (KDML) is concerned with the development and application of algorithms that can analyse data and derive useful information from it.
Find out more about KDML on our research pages.
Suggested research topics
Deep learning in medical image analysis
Supervisor: Yu-dong (Eugene) Zhang
Tumour segmentation is one of the most challenging tasks due to the unpredictable appearance and shape. Our team proposed a deep-learning based approach for segmentation of brain tumors, including tumor core and peritumoral edema. The AI-based delineation for tumors are comparable to or even better than human experts.
Detection of Alzheimer’s disease (AD) and mild cognitive impairment (MCI)
To differentiate MCI-stable from MCI-converter is important for following treatment. AI technique can help us do this task.
Multiple sclerosis (MS) Identification.
It is laborious to identify MS in early stage. AI has arrived a good accuracy in our study, and can identify MS patients from healthy controls automatically within seconds.
Breast cancer identification in digital mammogram.
The abnormal masses in mammogram appear white, so does the dense breast tissues. AI can help capture the neighboring information and help us recognise the abnormal masses from dense breast tissues in an accurate performance.
Cerebral microbleeding (CMB) in Susceptibility-weighted imaging (SWI).
SWI is more sensitive to detect microbleeding points. Our team proved deep learning can carry out CMB detection and following statistical analysis (area, volume, shape, etc.) within seconds.
Examine the development of basal-like breast tumours over time in histopathological images and to develop an automated system of analysing histopathological images for basal-like breast cancer detection.
Supervisor: Huiyu Zhou
Aims of this project are to examine the development of basal-like breast tumours over time in histopathological images and to develop an automated system of analysing histopathological images for basal-like breast cancer detection.
We will develop automated deep learning techniques to achieve robust pixel based and patch-based segmentation and classification (e.g. cancer vs. non-cancer regions). Using clinical histopathological images, we will experimentally demonstrate that the proposed basal-like tumour identification system at a system-level can produce more accurate and efficient results than the currently available systems reported in the literature.
Breast cancer is the second most common cause of death in women. The basal-like carcinoma is proposed subtype of breast cancer defined by its gene expression and protein expression profile. Breast cancer is divided into five subtypes, including luminal subtype A, luminal subtype B, normal breast-like subtype, HER-2 overexpression subtype and basal-like subtype. Basal-like tumours are one of five molecularly defined subclasses in breast cancer, and have drawn large attention of researchers because of its limited therapy opportunities. This tumour subtype occurs frequently in younger (<50) patients and constitutes approximately 15% of all breast cancers. Although basal-like tumors are characterized by distinctive morphologic, genetic, immunophenotypic, and clinical features, neither an accepted consensus on routine clinical identification and definition of this aggressive subtype of breast cancer nor a way of systematically classifying this complex group of tumors has been described. Therefore, we here propose a new deep learning based approach for improving the accuracy of basal-like breast cancer detection.
List of skills required:
- Ability to understand the design and analysis of algorithms.
- Ability to understand and apply basic concepts of machine learning and AI.
- Ability to present and write in English.
- Ability to write programs in Java, C++/C, Matlab or Python.
- Zhang, K., Crookes, D., Diamond, J., Fei, M., Wu, J., Zhang, P. and Zhou, H. "Multi-scale colorectal tumour segmentation using a novel coarse to fine strategy". Proc. Of BMVC, 2016.
- Moyes, A., Zhang, K., Crookes, D., Ji, M., Wang, L. and Zhou, H. "A novel method for unsupervised scanner-invariance using a dual-channel auto-encoder model". Proc. Of BMVC, 2018.
- See Huiyu Zhou's homepage and contact details.
Learning from failures
Supervisor: Mohammed Mousavi
Not all failed test cases are deemed significant in practice and many such failures are dismissed on the grounds that they are corner cases and are unlikely to surface in real practice. The goal of this project is to generalise single test case failures into classes of failed test cases and providing a symbolic representation of them.