Self-driving MRI

Chronic diseases such as hypertension, obesity and diabetes are rising steeply and increasingly individuals suffer from multiple conditions (comorbidities) - the treatment of the 15 million people in the UK accounts for 70% of inpatients in the NHS. The coexistence of several diseases makes diagnosis and treatment challenging, as they interact, hiding typical features or giving rise to additional pathologies. As a consequence, personalized diagnosis and treatment are becoming ever more important.

Similarly, pregnant women are increasingly affected by comorbidities, which contribute to major pregnancy complications. The success of care options such as planned delivery, medication and surgeries during pregnancy depends largely on personalized early diagnosis.

Fetal Magnetic Resonance Imaging (MRI) is a safe non-invasive imaging modality. It is perhaps the ultimate example where personalised examination is needed since both mother and baby routinely experience transient phenomena such as uterine contractions and sporadic exercising of the fetal chest muscles to train breathing.

MRI can produce images in any plane through the body with diverse contrasts reflecting different processes such as slow blood flow or oxygen concentration. Despite this flexibility, current fetal MRI examinations are surprisingly rigid in structure. The images are acquired in a predefined order, much of the time is spent inefficiently "chasing the baby" and repeating scans, all designed to present a static picture of an intrinsically dynamic situation. However, the way the baby moves, the frequency with which they 'breath' and the way they react to contractions can all provide information predictive of pregnancy outcome.

My project proposes a paradigm change, that turns MRI into a responsive, self-driving examination that embraces subject diversity and unpredictable processes to extract the most pertinent, eloquent and personalized information. Examinations in pregnancy are particularly challenging and feature unpredictable, but information rich, events. Instead of trying to make the baby behave in the way that is best for conventional MRI scans, I propose to put the MRI scanner into responsive mode. The whole womb will be assessed to automatically detect the position of the baby and any risk patterns such as previous C-section scars. The scanner will then be able to respond and adapt the scans. This is possible in real-time due to the recent development of 'machine learning (ML)', in which computer programs can be tuned to identify and respond to incoming information. ML requires lots of examples (in this case images with different anatomical features and behaviours) and fortunately I have access to a large data collection (>2000 fetal MRIs) from diverse pregnancies.

Furthermore, external sensor data (e.g. real-time Ultrasound) will allow the MRI to respond to important events such as contractions and fetal breathing while they happen. The entire MRI session will be changed from a series of blocks, which start and stop the flow of data, to a continuous responsive examination. We will deliberately include variation for a more comprehensive and personally focused examination and complement this with sophisticated analysis methods to quantify and present the gathered information.

My research will take place at King's College London, in a unique interdisciplinary environment with ML researchers, clinicians, and industry partners. After validation in healthy pregnancies, the methods I develop will be tested in 150 women with pregnancy complications and then put into clinical practise. At the end of this fellowship, this novel paradigm of "responsive self-driving MRI", in addition to providing richer indication-focused information, allow for the 1st time a comprehensive assessment and unique insight into the womb - crucial for best antenatal care. It aims to open up new avenues for personalized imaging well beyond the focused example.

The proposed research stream aims to contribute an innovative and potentially disruptive technological capability for MRI acquisitions. It addresses a growing need of the UK population by offering individualized imaging - suited for the rise in comorbidities, obesity and advanced age. It is carefully crafted and embedded into a clinical context and structured as an active collaboration between scientists, clinicians and industrial partners with a capability to directly test emerging innovations on patients. The focus on fetal MRI is well aligned with the UK's leading role in this field.
The planned ambitious program thus has the potential to contribute to a unique world-leading research activity.

There are three main pathways to my impact strategy:
1) Healthcare system, NHS, Patients and clinicians: Promote and raise acceptance for accelerated personalized imaging and cost reductions and raise acceptance. Of course, the primary beneficiaries will be patients, in the chosen application of fetal imaging specifically the pregnant women and their babies as well as the entire family affected. This applies for all pregnant women at risk of pregnancy complication, but specifically the growing number of pregnant women with comorbidities such as hypertension, diabetes or threatened of preterm birth. Achieving the aim to improve imaging accuracy and robustness will ultimately result in earlier, more comprehensive and personalized diagnosis, which facilitates targeted intervention, delivery planning and thus ultimately reduces morbidity, mortality and the risk for long term developmental disabilities associated with preterm birth. The availability and exploitation of additional information at the time of the scan reduces the need for repeated scans, thus shortens the time to diagnosis and subsequently cost and emotional stress. The proposed paradigm provides another by reducing costs - thus potentially contributing to decreasing the substantial and growing economic cost of the 3.4m MRI scans per annum (2017-2018) for the UK economy. Gains in efficiency would have significant impact on healthcare delivery and costs. Embedding of the project in the clinical setting and close collaboration with obstetricians caring for the identified high-risk population will shorten the pathway to clinical implementation and thus realization of these impacts.

2) Medical imaging equipment manufacturers: Promote collaborations to support clinical and industrial translation of advanced technology. A major barrier to translation for MRI methods is the need to translate them to the proprietary software run by each vendor's hardware. For maximal impact, inclusion of the developed technology and software into vendor product pipelines is thus crucial. Collaboration with two industrial partners will ensure this.

3) Academic beneficiaries: Promote collaborations to facilitate new research opportunities and innovative products. To assure the widest possible academic impact all my results and anonymized data will be publicly accessible, the results will be disseminated through open access publications, conference presentations and media (see statement "academic beneficiaries").To increase the potential for uptake of the methods we develop, we propose to disseminate the models as open-source software where appropriate and have included funding to access the previously successfully used xnat platform. All software developed, such as the pretrained AI networks, e.g. for the fetal biometry and risk factor detection, image reconstruction code, and fully anonymized datasets will be made available as open source through github (algorithms) and xnat (data).