Blog

What are your main research interests?

My key research interests are in various aspects of data science. For many years I have been mainly focused on analysis of visual information working in the areas of medical and industrial computer vision and machine learning. I am particularly interested in the use of Bayesian methodology for data modelling; pattern recognition; statistical shape analysis; deformation modelling for model-based recognition, segmentation, and registration. More recently I am also working on different applications of deep machine learning looking at classification, detection, segmentation, and depth estimation, but also expanding beyond standard applications of AI and developing deep surrogate models e.g., for inverse problem applications. By training, I am an engineer and therefore I am also interested in developing physical instruments, predominantly involving vision systems. Some of the most recent examples include smart mirror technologies for ubiquitous cardio-vascular self-monitoring, analysis of CT data for renal cancer immunotherapy treatment outcome prediction, early detection of sepsis, and decision support for colonoscopy procedures.

Can you tell me about the focus of your research?

A current significant focus of my own and my group research is finding ways of unlocking diagnostic potential of biomedical data with the objective to refine and find new biomarkers supporting diagnosis, treatment monitoring, and outcome prediction. A good example of such research is work we are currently doing on the “Machine Learning System for Decision Support and Computational Automation of Early Cancer Detection and Categorisation in Colonoscopy (AIdDeCo)” project funded by STFC CDN+. The project aims to develop new computational methods to improve robustness and effectiveness of colorectal cancer screening, with developments in areas of automatic detection, segmentation, and categorisation of polyps/lesions. We are also looking at developing computational aids supporting colon examination, including 3D structure visualisation as well as navigation tools, which could be particularly important when the colonoscopy is used in conjunction with other techniques such as capsule-colonoscopy and/or CT-Colonoscopy. Capsule-colonoscopy is a relatively new technique for colon examination, where a capsule equipped with cameras automatically records images during its passage through the digestive tract after being swallowed. Although it is less intrusive than colonoscopy it has its own challenges, limitations, and disadvantages. Machine learning can be instrumental in eliminating some of them, e.g., automatically highlighting parts of the recorded video (lasting more than an hour) for examination by a clinician.

AI and machine learning has the potential to significantly affect how healthcare is delivered. Certainly, the intensity of research and number of commercial products entering the market seem to support such a proposition. Some of the potential applications may not be this obvious and indeed don’t have to directly involve analysis of data. For example, one possible AI application we are investigating in the context of the AIdDeCo project is automated report writing, which would not only free clinicians’ time from a routine time-consuming task, but also would provide a more objective record of the procedure with key actions automatically documented, in a way providing an analogue of a black box used in aviation.

What are the potential impacts this research may have?

AIdDeCo is one of the projects that have a potential to have a tangible impact. The research is important as colorectal cancer (CRC) is one of the leading causes of cancer deaths worldwide, with the survival rate depending strongly on an early detection, hence the importance of the colon screening. Colonoscopy is widely accepted as the gold standard for colon screening due to its ability of detecting and treating the polyps/lesions during the same procedure. However, colonoscopy has some limitations. Recent studies have reported that about a quarter of colon polyps (adenomatous polyps are commonly accepted as precursors for colorectal cancers) could be missed during routine colonoscopy screening procedures, with a sizable number of patients having at least one polyp missed. It has been also estimated that improvement of polyp detection rate by 1% reduces the risk of CRC by 3%. It is therefore essential to improve polyp detectability and, in general, to develop methods supporting endoscopists during colon examination procedures. The AIdDeCo project is aiming to do exactly this by automated video analysis leveraging recent advances in machine leaning. The key output of the project is a set of software tools aimed at improving polyp detectability and therefore ultimately leading to a reduced risk of colorectal cancer.

Building on the existing base of UK and EU collaborators, the AIdDeCo ambition is to create a strong multifaceted international consortium capable of developing, clinically validating, and deploying automated colonoscopy systems to clinical practice. It is envisaged that the project will be instrumental in building such a consortium, which would subsequently apply for funding to support further research and secure the transfer of the developed technology to a higher Technology Readiness Level.

What has been the highlight of your research career so far? 

It is not as simple to answer such question as it may seem. We tend to appreciate our most recent achievements the most. However, one of the most enjoyable aspects of my work over the years were my interactions with my Post Docs and PhD students. Following their professional success is a constant source of enjoyment.

There is also a buzz when ideas, methods, or tools you helped to develop are used by others. For example, I have been a part of a European project doing some early work on methods and technologies for contactless measurements and interfaces for smart mirrors. Recently there seem to be numerous devices which have moved these ideas into the commercial space. Other examples include software popular with users or highly cited papers.

And what outcomes of your research are you most looking forward?

There is of course the already mentioned AIdDeCo project, which I am passionate about and hope that the work we do on that project will find a way to clinical practice.

Recently, I have also started collaborating with physicists and mathematicians on developing deep machine learning surrogate models. These are still early days for me, but I am really excited about the prospects these methods can bring to some advanced computational methodologies, which otherwise would not be used in practical applications due to prohibitive computational complexity.

What advice would you give to someone considering a career in scientific research? What are the highs and lows so far?

A career in research can be very rewarding, but it takes a lot of work and perseverance. You should expect lows when your ideas / papers are rejected and there are times when it is difficult to get support for your research. However, these lows are more than compensated by the highs when things start working, especially after prolonged effort when success seemed all but impossible.

Research Interests? 

My main interests are multimodal gamma ray cameras for medical and non-clinical applications, along with novel X-ray detectors. My research interests have expanded into imaging technologies for nuclear security and post clean-up operations that build on the hybrid camera concept. This has led to collaborations with the National Nuclear Laboratory, Sellafield Ltd and industry partnerships. Recently, I have taken an interest in Perovskites, a group of materials that offer high efficiency X-ray detectors and scintillators. I am part of a new project, led by Dr Sam Stranks, at University of Cambridge to explore these new materials by developing highly sensitive X-ray detectors using halide perovskite (PVK) semiconductors - materials currently making impact as disruptive photovoltaic (PV) technologies - for phase contrast X-ray imaging. A key application is for medical imaging through low-dose, high-resolution and inexpensive computer tomography (CT) scanners where highly innovative hardware and software components will be developed side-by-side to enable automated all-in-one pre-symptomatic diagnosis. 

The focus of your research? 

High spatial resolution imaging systems with good efficiency and multimodal functionality have the potential to improve healthcare in several key areas, such as cancer diagnostics and accessibility of bedside diagnostics or remote care. Understanding the healthcare need and providing affordable solutions is a key driver for my research. 

What are the potential impacts this research may have?

The portable hybrid gamma camera offers healthcare benefits by providing patients and medical teams with versatile, point-of-care technology, and represents a step-change in such imaging. Following ethics approval, the camera was evaluated in clinical imaging trials at the Queens Medical Centre, Nottingham. The success of these trials led to the technology being licenced to Serac Imaging Systems Ltd who are now developing a commercial imaging system for the healthcare market. Looking to the future, highly sensitive X-ray detectors using halide perovskite (PVK) semiconductors in conjunction with AI-driven algorithms for image reconstruction, lesion detection and segmentation will realise quicker and more efficient healthcare delivery and prevent disease spread through extremely early detection and for routine follow-up of oncology patients.

What made you decide to have a career in physics?

I have always had an interest in physics, watching programmes such as Tomorrows World, the Feynman Lectures and Carl Sagan’s Cosmos. However, I left school when I was 16 years old to become an apprentice electrician. A few years later my younger brother, who also left school at 16, returned to higher education to study physics and he inspired me to begin an undergraduate degree in Physics at the ripe old age of 23.

What has been the highlight of your research career so far? And what outcomes of your research are you most looking forward?

There have been a number of highlights. My earlier research work was on X-ray imaging systems for satellite missions, and I was part of the Chandra X-ray Observatory that was successfully launched in 1999. This observatory is seen as the X-ray equivalent of the Hubble Space Telescope. Knowing that my research improved one of the imaging cameras (HRC) on Chandra that has been the basis of many outstanding astrophysical discoveries is something that I am very proud of. Another highlight is the development of the hybrid gamma camera that will soon be available to surgeons and clinicians around the world to help them diagnose and improve treatment of a range of cancers.

What advice would you give to someone considering a career in scientific research? What are the highs and lows so far?

Try to find an area that is exciting to you but be aware that research can be frustrating and hard work. Also be realistic and grab opportunities when they arise. Getting a full-time academic post is still difficult and other roles in industry should always be considered. A low was being rejected for an academic post when I thought my experience fitted the required role extremely well. Highlights, includes forming my own research group and helping all my postgraduate students to successfully complete their PhDs.

At their core, both problems seek to overcome the same challenge: multi-scale complexity'

For both cosmology and cancer research the challenge faced utilizing big data is fundamentally the same. The scientist has numerous observations, which each provide a narrow snapshot into a specific aspect of the overall system. Their goal is to extract meaningful information from an extremely large dataset, using robust and reliable statistical techniques.

For example, in my Cosmology research, I analyse images of tens of thousands of galaxies. Each image contains an extremely small amount of information about the dark matter contents of the region of the Universe that the galaxy resides. However, extracting this signal requires that I separate it from the intrinsic properties of the galaxy itself, necessitating detailed modeling and analysis of every individual dataset.

Healthcare researchers working in big data are faced with a similar task, albeit their data is now medical in nature and they wish to separate out what each observation tells them about the treatment of individual patients from global trends in patient outcomes. However, the statistical methods underpinning their analysis are the same ones I use to study dark matter, they are simply being applied in another scientific domain and to a completely different problem!

A couple of years ago, I by chance had lunch with the CTO of Concr, a Biotech company who make predictive models of how cancer therapy may work. We got chatting about statistics, Bayesian inference, model-fitting -- and quickly realised we had a lot more in common than we initially expected! 

We began working together on a project we were soon referring to as “Cosmology & Cancer”, with our aim to collaborate on the development of statistical methods that could overcome the challenges we both faced in the big data era. With the healthcare company Roche and the Christie NHS Foundation Trust we are now working on an Innovate UK clinical trial. This is centred around Carcinoma of Unknown Primary (CUP) Site, where malignant cancer cells are found in a patient but the original cancer is unknown. This makes it extremely challenging to treat, withcurrently no approved therapies or immunotherapies available in the UK.

With Concr, I oversee development of a multi-level modeling framework to fit their models of cancer to clinical data on CUP patients. This centres around the shared development of open-source statistics software titled PyAutoFit (https://github.com/rhayes777/PyAutoFit), which hosts the inference techniques we both need to analyse large datasets.

For CUP diagnosis, we are building multi-level graphical models of cancer. These models aim to properly account for multi-scale complexity, by hierarchically fitting datasets that provide information at all relevant physical scales. The inner levels comprise data on the cellular and molecular scales (e.g. genetic or epigenetic profiles) to infer how different cancer cells respond to treatments. Higher levels combine these models to represent the evolution and dynamics of tumours, whilst the highest levels map out effective treatment paths using data on patient outcomes.

By comparing these results to known data of a CUP patient, the hope is to build a predictive model that can determine the original cancer and provide an evidence-based pathway for treatment.

The project has now come full circle, with this framework finding its way into my Cosmology research. The inner levels are now models on the scale of individual galaxies, whilst the higher levels describe perhaps the largest scale conceivable -- that of the entire observable Universe. Remarkably, code that someone else wrote to improve cancer therapy is helping me better understand our Universe!

When we set out on this project, our goal was to simply share code for analysing large datasets. However, as I become more immersed in biological research, I am starting to question if there is a more fundamental link with Cosmology. At their core, both problems seek to overcome the same challenge: multi-scale complexity. As I study them further, symmetries are emerging in my understanding of both problems, and it is uncovering whether this link is more than coincidental that motivates me to pursue this research further.

However, it is in my personal and professional development level this project has had the most profound impact. I have been taken aback by how much I have learned working with those in other scientific disciplines. It has been an eye-opening and incredibly positive experience from which I have learnt numerous statistical techniques, software development practises and approaches to “doing science” that I was simply not exposed to in Astronomy. Adopting these has furthered my own cosmology research and grown me as a scientist, and I am passionate to share my experience with the wider field of Astronomy.

What are your main research interests?

My research relies on X-ray spectroscopy, and so one of my main interests is to improve analysis by making corrections to the spectra based on the underlying physics. A lot of this work is based upon theory and modelling the problems with computer simulations, and I enjoy this challenge. It is great however that I get to combine this with taking my own measurements in the lab. In the future, I look forward to being able to test some of these algorithms by applying them to measurements of some phantoms. 

What is the focus of your research? 

The goal of my research is to create a method to measure the density of breast tissue. Breast tissue can be broken down into adipose and glandular tissues, the proportions of which govern its density. By examining the exposure of such a tissue to X-rays, it is possible to reconstruct how much of each constituent tissue there is, and therefore calculate the density. 

What are the potential impacts of this research?

Studies have found a link between breast density and breast cancer risk, so such a measurement could form part of the screening process to improve its efficiency. 

What made you decide on a career in physics?

At each educational stage, I’ve always picked the subjects that I enjoy the most, which led me to studying a degree in physics. I chose to go into research over other routes because I love the idea of finding some new information that hasn’t been uncovered before. Not every discovery can be ground-breaking, but it would be amazing if I could expand human knowledge by a tiny amount. 

What has been the highlight of your research career so far?

So far it has been my transition into working in the lab. I love the area of physics I’m working in, but previously I did not have any practical experience of taking my own measurements using X-rays. I’ve really enjoyed having the opportunity to apply the theory within the lab. 

What advice would you give to someone considering a career in scientific research? What are the highs and lows so far? 

There will be times you explore an alley that later closes, and I think it’s important not to be put off by this. I think this is outweighed by the freedom to explore an area you really enjoy. For anyone who is considering a career in research, I would say go for it! 

“Improvement of polyp detection rate by 1% is estimated to reduces the risk of colorectal cancer by 3%. With the help of new machine learning methodologies, it seems conceivable to significantly increase robustness and effectiveness of colorectal cancer screening improving polyp detection rate.”

STFC CDN+ is supporting a project aiming to develop new computational methods to improve robustness and effectiveness of colorectal cancer screening, improving lesion detectability and provide endoscopists with practical navigation tools, in effect reducing cost, risk and discomfort to patients.

Colorectal cancer (CRC) is one of the leading causes of cancer deaths worldwide, with the survival rate depending strongly on an early detection, hence the importance of the colon screening. It is commonly accepted that most colorectal cancers evolve from precursor adenomatous polyps. Typically, a colonoscopy screening is proposed to detect polyps before any malignant transformation, or at an early cancer stage. The optical colonoscopy is the gold standard for colon screening due to its ability of detecting and treating the lesions during the same procedure. However, colonoscopy screening has some limitations.

Various recent studies have reported that about a quarter of colon polyps could be missed during routine colonoscopy screening procedures, with a sizable number of patients having at least one polyp missed. It has been also estimated that improvement of polyp detection rate by 1% reduces the risk of CRC by 3%. It is therefore essential to improve polyp detectability as it can reduce the risk of colorectal cancer and healthcare costs. Equally, correct classification of detected polyps is limited by polyp appearance and subjectivity of the assessment.

“My work on the analysis of video colonoscopy data started in 2015 with participation in the Endoscopic Vision Challenge organised as part of the Medical Image Computing and Computer Assisted Interventions (MICCAI) conference. This was done following EPSRC-funded studies that I had been conducting since 2007, in collaboration with medical physicists and optical engineers, and focused on the applications of computer vision in radiotherapy. Teams from my Computer Vision and Machine Learning (CVML) research group continued to participate in the subsequent Endoscopic Vision Challenges, which we won twice. Through these years we have built up an expertise in this subject area, reflected by the successful completion of a PhD project, and managed to establish collaborative links with UK and EU academic and clinical partners. The STFC CND+ proof of concept funding is facilitating collaboration with computational physicists to further optimise, develop and integrate existing solution to a stage that it can be independently assessed by our clinical partners, with a view of taking the created software to a Technology Readiness Level (TRL) suitable for clinical study/trial”

 The key challenges and expected project innovations are to advance the current state-of-the art in deep machine learning applied to colonoscopy data. The focus is on developing techniques for automatic detection, segmentation and categorisation of polyps, aiming to reduce chances of polyps being missed during colonoscopy procedures. Furthermore, we aim to detect image artefacts (e.g. saturation, low contrast areas, blur or specularity) and foreign objects to support endoscopists in interpreting the data during examination. We are also looking at developing visual aids supporting colon examination. These will include 3D structure visualisation as well as navigation tools based on visual odometry techniques.

The key expected output of the project is a set of software tools aimed at improving polyp detectability and therefore ultimately leading to a reduced risk of colorectal cancer. The developed software is to be tested and evaluated by clinical endoscopists, preparing grounds for future clinical studies validating the methodology in a hospital environment.

To take the full advantage of the available resources and the research effort, we will also consider secondary application areas to fully utilise the resources and the project outcomes. We intend to test the developed machine learning methods on different medical applications. In particular, through transfer learning, we expect to validate developed machine learning solutions on a radiomics problem, quantifying CT scan density heterogeneity in renal cancer.

Building on the existing base of UK and EU collaborators, it is expected that the project will lead to creation of a strong multifaceted international consortium capable of developing, clinically validating, and deploying automated colonoscopy systems to clinical practice. It is envisaged that the project will be instrumental in building such a consortium, which would subsequently apply for funding to support further research and secure the transfer of the developed technology to a higher Technology Readiness Level.

“An important motivation for my work, including this project, is to see how state-of-the-art computational methods can be translated into tools that can have a real impact on the health and quality of life for us all.”