AI for Population and Global Health in Radiology

Udunna C. Anazodo, PhD, is an assistant professor of neurology and neurosurgery at the Montreal Neurological Institute at McGill University. She is the founder and chair of the Consortium for Advancement of MRI Education and Research in Africa (CAMERA) and is currently leading efforts to create the Africa Neuroimaging Archive (AfNiA). Her research interests include diagnostic image analysis using artificial intelligence methods to enable quantitative PET and MRI for population neuroscience and global health.

Maruf Adewole, MSc, is a medical physicist. He holds a bachelor's degree in physics and master's degree in medical physics from the Federal University of Technology Akure and University of Lagos, Nigeria, respectively. He is currently a PhD student in the department of radiation biology, radiotherapy, and radiodiagnosis of the University of Lagos and the lead artificial intelligence scientist at the Medical Artificial Intelligence Laboratory at the University of Lagos. His doctoral project is focused on the use of artificial intelligence in the segmentation of brain tumor for low-resourced settings.

Farouk Dako, MD, MPH, is an assistant professor of radiology in the division of cardiothoracic imaging, director of radiology global and population health research in the department of radiology, and scholar in the Center for Global Health at the University of Pennsylvania. He is also the director of the RAD-AID Nigeria program.

While the debate regarding the value of artificial intelligence (AI) in radiology rages on, the World Health Organization (WHO) approved the use of AI in place of human readers to detect tuberculosis (TB) on chest radiographs in its updated TB screening guidelines and recommendations (1). This move is in keeping with the growing recognition that the potential for AI to improve population health outcomes is likely higher in low- and middle-income countries (LMICs) due to their greater disease burden and the limited availability of the needed health care infrastructure and personnel. Limitations in infrastructure also present challenges in the implementation of AI solutions, which often require prerequisites such as a picture archiving and communication system (PACS), internet connectivity, and reliable power supply. Infrastructure to support the creation of local disease databases is important for ensuring AI solutions are tailored to be valid in the health care environment in which they are deployed. Data diversity also enhances generalizability and reduces the bias of AI algorithms and is critical to preventing the exacerbation of disparities in health outcomes that exist along demographic and geographic lines. Policy to govern data storage and sharing is needed to ensure patient privacy and standardized procedures and to enable equitable relationships between data owners and users. The advancement of AI in global and population health requires approaches to improve infrastructure, data quality, and governing policies.

Before implementing AI solutions, analyzing the local health care landscape is of utmost importance in order to understand the health care needs; to identify the presence, location, and owners of resources; and to map out socioeconomic factors and cultural and political drivers. AI solutions should be driven by population health care needs and should be sensitive to local factors (2). For example, according to the WHO, the three leading causes of mortality in LMICs before the COVID-19 pandemic were ischemic heart disease, stroke, and lower respiratory infections. This information should be incorporated strategically to prioritize the development and deployment of AI solutions into these regions. Efforts carried out in collaboration with local partners designed to increase local capacity are more likely to be effective, scalable, and sustainable. RAD-AID International has been successful in collaborating with industry and academic partners to advance health information technology and AI in Nigeria and other LMICs (3). Following a needs assessment, a hybrid cloud-based PACS system was installed at University College Hospital, Ibadan, Nigeria. This system allowed for the subsequent donation and installation of AI software for chest radiograph interpretation that has undergone a local validation process and has been incorporated into the clinical workflow (3).

For LMICs to benefit from global and population health AI solutions, models developed must be capable of producing imaging biomarkers that enable country-specific and population-level assessments of relevant disease traits and trends. It is unclear if state-of-the-art AI solutions developed largely in the Global North can be implemented widely in LMICs, given the extensive presence of older-generation imaging technology in LMICs. For example, a substantial proportion of brain MRI studies acquired in clinical settings across sub-Saharan Africa (SSA) have poor image contrast, low resolution, and inherent noise. Although AI solutions exist to enhance image quality (4), the lack of annotated imaging data from SSA further limits the implementation of state-of-the-art AI models. To provide the enabling environment for the generalizability of AI solutions in SSA, we are creating the Africa Neuroimaging Archive (AfNiA), a publicly available imaging repository of annotated brain MRI studies. AfNiA will not only aggregate clinical brain MRI data linked to ground truth labels, but, more importantly, will provide a backbone to solve long-standing infrastructure, data quality, policy, and capacity challenges that bar the inclusion of LMICs in global disease diagnosis and prediction efforts. In collaboration with local partners, a comprehensive platform is being developed for data preparation compliant with FAIR principles (Findability, Accessibility, Interoperability, and Reusability) (5) and conventional data privacy and protection regulations to streamline the series of time-consuming data processing steps, from de-identification of personal information to data quality verification and creation of labels for AI model validation and testing. Policies for ethical use are being developed in lieu of region-specific data protection provisions to govern access to annotated data and ensure that AI solutions created using AfNiA will directly benefit the local communities who own and supply the data. A process for data transfer from a network of local clinics to the AfNiA repository is being implemented to enable participation from centers without PACS or internet connectivity, which are often the least-resourced and more remote settings in the region. The first AfNiA data release will support global efforts to improve brain tumor classification, treatment, and survival prediction. Supported by the Lacuna Fund in Equity & Health, we have partnered with the Brain Tumor Segmentation (ie, BraTS) Challenge (6) to enrich the challenge data and evaluate state-of-the-art models for glioma tumor segmentation and classification (7). Future release of annotated pituitary adenoma, ischemic stroke, and epilepsy brain MRI data are planned.

To implement and sustain AI solutions for global and population health, local capacity in regional AI ecosystems must be established for data collection, model construction, and clinical translation. The Lacuna Fund will support efforts to train African AI communities in computational imaging processes through a regional neuroimaging informatics hackathon tied to the public release of the AfNiA brain tumor data. This capacity-building effort will enhance local skills in AI data curation and model development, as well as incentivize participation in grand challenges.

In general, AI has the capacity to reduce global health disparities through computer-aided diagnosis and identification of biomarkers for population-based prevention solutions. As a global community, we have succeeded in controlling disease mortalities simply by understanding disease causes and addressing common modifiable risk factors. Our recent collective effort to understand the SARS-CoV-2 viral genome and epidemiology contributed to early detection and primary prevention strategies that reduced COVID-19 mortality globally, even in the absence of effective vaccines or therapies (8). Likewise, to solve prevailing global and population health challenges, improvements in infrastructure, data quality, policy, and capacity across all settings are paramount.