MICA: InterdisciPlInary Collaboration for efficienT and effective Use of clinical images in big data health care RESearch: PICTURES

Clinical imaging including X-rays, CT, MRI, ultrasound and nuclear medicine scans are core diagnostic technologies. These images can support many important areas of research to improve any or all of diagnosis, monitoring of disease progression and response to treatment. Currently most research using images is based on those collected specifically for a particular research project. The images are of "research" quality. That means that they are captured at high resolution using standardised procedures to reduce the variability of images. Research data collection is expensive so studies tend to be small, and the people who take part in research studies are different to those seen in normal clinical care. It is therefore often uncertain whether research findings can be translated to "real world" images or patients.

Each year millions of clinical images are generated in Scotland through routine examinations at hospitals and stored in a huge database. The Scottish national imaging database currently has ~23 million different images collected since 2010. Access to these "real world" images would be extremely valuable for research, but there are a number of big challenges. Firstly, it is very important that all data is kept confidential. Secondly, imaging datasets are very large which it technically challenging. Thirdly, the software which manages these images is optimised for retrieval of images by NHS staff specifically for an individual patient's clinical care (e.g. return Mrs Jones' scan taken on the 28th of May 2016) rather than for research (e.g. return all the CT chest scans of smokers between age 55 and 65 where a contrast agent has been used).

What will be delivered? This 5 year programme will enable secure access to routinely collected imaging data for research. Using the foundation blocks already in place from previous research grants, PICTURES will extend, scale and enhance innovative open source software to query a research copy of the Scottish National imaging database securely hosted by the University of Edinburgh and provide anonymised extracts of hundreds of thousands of images for research. PICTURES will also develop this software to query imaging data linked to genomic data securely hosted by the University of Dundee.

There are 3 main areas of research required within the core programme: (1) Data science research for complex cohort building from real-world, messy data. (2) Engineering required for scaling and handling big data within a Safe Haven environment. (3) Cybersecurity research needed to ensure that the patient data is securely held and de-identified appropriately for research.

PICTURES will support 2 major exemplar research projects to guide and shape the underpinning resources. Exemplar one will develop a method to detect lung nodules and coronary artery calcification using hundreds of thousands of CT chest scans provided by the core programme. It will also predict the risk of getting lung cancer based upon the presence of lung nodules and the risk of cardiovascular disease based upon the presence of coronary artery calcification. This exemplar will work in partnership with an industrial partner, Aidence, to validate and test the method directly in NHS clinical workstations within the course of the programme.
Exemplar 2 will predict individual risk of dementia in people with diabetes using MRI brain scans, genetic data and medical records. The most important variables will be found. The predictive tool will be validated on the large image dataset provided by the core programme.

Both of our exemplars will determine new information from routinely collected data that would otherwise have been ignored. Predicting and therefore treating diseases at an early stage improves patient outcomes and reduces the cost to the NHS.

PICTURES is truly interdisciplinary requiring expertise in Radiomics, AI, Cybersecurity, Software Engineering, Data Science, Data Governance and Medicine.

The core programme focuses on 3 key areas:
(1) Complex cohort building from noisy, heterogeneous data: we will develop algorithms for text mining and standardising imaging metadata. We will use machine learning classification techniques to group images and natural language processing to pre-process and mine knowledge features from the free-text data and remove the identifiable data. We will utilise imaging processing algorithms to search for features within the core dataset to build new cohorts based upon pixel data.
(2) Scaling and handling big data: we will develop algorithms and optimisation processes for handling petabytes of imaging data and investigate how to optimise the use of GPUs within a virtual environment.
(3) Cybersecurity: we will ensure that our systems are secure but also meet the requirements of the research community.

Exemplar 1: Using CT scans provisioned from the national resource, we will use Deep Learning to train an algorithm to detect lung nodules and coronary artery calcification. The algorithm will be implemented as a Medical Device and made available within NHS Clinical PACS reporting workstations for Clinical performance evaluation and validation. To determine the risk of lung cancer and cardiovascular events we will use the nationally available longitudinal health outcomes data linked to the results from the CT scans.

Exemplar 2: To develop a risk score for dementia and an understanding of which data is most important for prediction, we will use voxel based feature selection and support vector machines within a cross-validation framework for MRI brain images. Non-imaging analyses will be primarily cross-sectional and longitudinal including time variable exposure clinical covariates as well as genomic covariates. Risk predictions using image, genetics and clinical data for individual patients will be combined using decision tree methods to create a best overall predictor. Algorithm validation will use the national resource.

Industry: There are a range of companies who will benefit from access to the imaging data e.g. Imaging Equipment Companies (Canon, GE, Siemens, and Philips), Imaging Contrast Media Companies (GE, Guerbet, Bracco) and Medical and Surgical Device Companies (Medtronic, Baxter).

Costs for the academic community: Scalable access to large quantities of routinely collected de-identified images via automated, reproducible processes will reduce the effort of obtaining the data required to answer research questions at scale. These resources will also reduce the effort of obtaining governance.

Widening and optimising access to data: Access to such large numbers of real world images has previously been very challenging. These resources should accelerate research in the field.

Patients: Increasing the availability of large scale routinely collected images linked to other forms of health data for both industry and academic use will lead to a greater likelihood of achieving results translatable into diagnoses and treatments.

Policy makers: There are many advantages of using a Safe Haven model for access to sensitive data. The models developed on the safe handling of big data for research may become an example of good practice worldwide, further raising the profile of UK healthcare research.

HDRUK: The UK wishes to be an internationally recognised centre for population based data research. These resources will add to the complement of excellent data available internationally.

Capacity Building and training: This programme will train the team in data science. There is a recognised shortage of expertise in this field (see Life Sciences Industrial Strategy and the ABPI Bridging the Skills Gap report).

Supporting Learning Healthcare Systems: This model has become a widely recognised approach applying a key principle of a feedback loop from the outcomes of research to directly improve clinical care. The work within PICTURES strongly supports this model.

Both exemplars will generate valuable knowledge in the use of routine clinical imaging as a source of potential clinically useful biomarkers. This will have a major impact on the cost effectiveness of clinical imaging in the NHS as potentially more clinically relevant information will be extracted beyond the clinical indication for the image.

Exemplar 1: We anticipate that the existing lung nodule detection tool will seamlessly interact with the coronary artery calcification tool, which will be developed. This will be relatively straightforward to integrate into PACS reporting workstations, allowing for immediate reporting of these findings during chest CT studies. Examples already exist of this and similar software tools, integrated within a variety of PACS vendor workstations. Once this integration is rolled out throughout the NHS, this should significantly improve the efficiency of radiology reporting, reduce errors of reporting and allow better individualised risk profiling and management options for patient care.

Exemplar 2: There are 3 significant impacts: (1) Enabling the identification of diabetics at risk of developing dementia affording clinicians the opportunity of optimising preventative measures. (2) Potentially facilitating the identification and recruitment of high risk individuals prior to dementia onset into clinical trials of prevention, an urgent need in dementia where trials of treating dementia once it has occurred have not been successful. This process will be augmented by integrating with the SHARE programme in Scotland. (3) The use of feature selection techniques will allow the identification of biological targets as disease biomarkers and/or therapeutic surrogate targets, this is in contrast to deep learning approaches whereby the discrimination is agnostic to biological structure or function.