Enabling Clinical Decisions From Low-power MRI In Developing Nations Through Image Quality Transfer

The long-term vision motivating this project is of software solutions that enable low-power cheap-and-sustainable imaging devices able to provide point-of-care image data in resource-poor locations at diagnostic/prognostic quality. We achieve this by propagating information from databases of high quality images. We provide a proof of concept using MRI from lower-power scanners available in LMICs, specifically Nigeria, that we enhance by propagating information from databases of images from state-of-the-art MRI scanners available in the UK. We focus on an application to childhood epilepsy to demonstrate early clinical benefit. Childhood epilepsy presents an immediate clinical need in LMICs, as MRI from widely available 0.36T scanners is insufficient to support clinical decisions on curative surgery that are routinely made in the UK using 1.5T or 3T images. This leaves many patients untreated, living with severe epilepsy and resulting physical disabilities and mental disorders, unable to work effectively, and draining sparse medical and social-care resources.

We draw on the latest advances in machine learning to approximate the MRIs available in the UK from those accessible in the paediatric neurology clinic in UCH Ibadan, Nigeria - a typical sub-Saharan city hospital. Machine learning has made major advances over the last few years. In particular, it shows remarkable feats of artificial intelligence in data-rich application areas such as computer vision where, for example, computers now outperform humans in object recognition. The advances are just starting to make an impact in medical imaging, which presents unique challenges because a) less data is available than many non-medical computer vision tasks, b) decisions are often more critical as they impact directly on patient outcome.

Our recent image quality transfer (IQT) framework propagates information from high quality to low quality medical images. It shows compelling early results, such as revealing thin white matter pathways, usually only accessible from specialist high resolution data sets, from standard resolution images acquired on a clinical scanner. Here we advance IQT to exploit the latest machine learning techniques, enhance those techniques to provide confidence measures valuable for medical decision-making, and tailor solutions specifically to enhance images from the Ibadan paediatric clinic with those from similar cohorts in the UK. We acquire and collate the data sets sufficient to support learning the required image-to-image mappings. Matched pairs of images from the same subjects from UK and Nigerian scanners are not practical to obtain, so we employ unsupervised and semi-supervised learning to construct image-to-image mappings without directly matching training data. We refine promising implementations and assess their impact on clinical decision making in a pilot study in Ibadan using locally agreed metrics.

We intend this project as a springboard for a much wider and long term program exploring these ideas to bring about a paradigm shift in imaging that deploys cheap point-of-care devices built specifically to acquire data enhanced by databases of high quality images acquired on state of the art or bespoke devices.

Impact comes through access to information from state-of-the-art image quality in locations where it has previously been unavailable. Although state-of-the-art MRI technology is synonymous with effective diagnosis and treatment in many medical specialties, developing nations (80% of world's population) still rely on older technologies such as low-field MRI that are insufficient for diagnosis and care management. Newer MRI technologies have often prohibitive purchase and running costs. Even more prohibitive, however, is the unreliable electricity supply and lack of availability of coolants necessary to keep the equipment operational, since state-of-the-art MRI equipment must remain permanently operational. Our project provides a low-cost, sustainable, and robust software solution based on machine learning and quantitative image processing.

The most direct impact is on paediatric epilepsy diagnosis and sub-sequent management in sub-Saharan Africa (Nigeria), which we use as a demonstrator for the new technology within the project. Epilepsy is the leading childhood neurological disorder worldwide, with 90% of the developmental socio-economic burden borne by resource-poor nations. Early diagnosis of focal brain abnormalities through localisation of epileptogenic lesions, associated with medically intractable epilepsy, helps identify patients for definitive surgery, which has high probability of being curative. This is particularly effective for children and recurrent patients, whose lives improve dramatically by cessation of seizures, which limits disrupted neuro-development.

In Nigeria, which lacks private health insurance schemes or a national health system, clinical epilepsy management accounts for up to 20% the annual income for families. Misdiagnosis from lack of quality MRI affects childhood development and quality-of-life and prolongs healthcare costs (diagnosis and medicines), direct non-medical costs (transportation and care) and indirect costs (lifestyle changes). Early diagnosis and appropriate treatment alleviates the strain on a sparse and over-stretched healthcare infrastructure by reducing clinical and care demands. Additionally, it reduces the socio-economic stress on the affected family by relieving behavioural problems associated with epilepsy diagnosis enabling parents to return to work and resume normal lifestyles. These impacts extend well beyond Nigeria to LMICs in Africa, Asia and South America where situations are similar.

Longer term, the same technology can make impact in a much wider range of clinical situations in which MRI is used routinely in the UK (and other higher income countries): psychiatric disorders, other neurological disorders, cancers (brain and non-brain), orthopaedics, respiratory diseases, and many others. Moreover, the technology extends to other imaging modalities. For example, CT is a common diagnostic modality across developing nations; similar techniques can enhance image quality and downstream decision making.

Ultimately, we envision a new paradigm of imaging built around the ability to enhance sparse point-of-care data with centralised databases of high quality data via machine learning. Cheap, portable, low-power devices will be manufactured specifically to exploit this ability bringing the power of state-of-the-art imaging even to the most remote and resource-poor locations and situations.