Hi, I'm Brijesh.

This was my Master's project conducted at the University of Manchester with collaboration with The Christie NHS Foundation Trust. The first part involved setting up a project pipeline to read in MRI image files (in NIfTI file format), pre-process them and storing these images. These were then used for large feature extraction using a package called radiomics.

The second part involved using the above pipeline, and a U-Net style convolutional neural network (CNN) to employ deep learning; this was used to segment two main defined cancerous regions - the gross tumour volume (GTV) which indicates the visible, macroscopic spread, and clinical tumour volume (CTV), which includes the GTV plus some of the surrounding area (containing possible microscopic spread).

GTV segmentation had also never done before using deep learning for cervical cancer at the time. The technique employed transfer learning using the ImageNet database (due to the limited size of the dataset of cervical cancer images) and in the end, accuracies of 55-80% were obtained in identifying the correct GTV/CTV regions. This project was predominantly Python-based though some MATLAB and Lua code had to be written as well. For future work, it would also be possible to repeat this technique on other sets of clinical images.

This project involved reviewing techniques in generating synthetic data, driven by the need of sharing sensitive data for R&D purposes. The synthetic data generation tools aimed to mimic characteristics of the real dataset whilst removing or obscuring the sensitive information present within the real data. The final report was made public here on Gov.UK.

For example, we might want the synthetic data to retain the range of values of the original data with similar (but not the same) outliers. Or we might want to retain a similar frequency distribution in the synthetic and original datasets. However, this becomes more complex when we start to consider interactions between fields, or different types of data such as free text and GPS locations.

Working in a team, I helped explore public datasets which were similar to those used by the customer and identified tools which could be used to produce synthetic data. For example, the Czech financial database comprised of numerical/categorical data was used with the Synthetic Data Vault (sdv) Python package whilst the text-based resumes/CSVs were used with Microsoft's open-source tool called "Presidio" (which was also powered by AI algorithms).

SDV worked with relational tables and from the testing and analysi using cross-correlation heat maps and histograms, the tool produced mixed results since it only seemed to work with Gaussian-distirbuted data. This often led to odd results such as negative bank ID values - however, since the tool is open-source, it's still currently in development and the ability to specify other distributions may be added in at a later date.

The final report can be viewed here on Gov.UK here, whilst a blog article regarding this investigation can be found here.