Content uploaded by Udunna Anazodo
Author content
All content in this area was uploaded by Udunna Anazodo on Aug 15, 2022
Content may be subject to copyright.
MICCAI 2022 FAIR WORKSHOP
The African Neuroimaging Archive (AfNiA): An
Approach for Democratizing AI in Medical Imaging.
Udunna C Anazodo1,2,3, Raymond Confidence2,3, Dong Zhang2, Abiodun Fatade4,
Johnes Obungoloch5, Can Akgun6, Aaron Mintz6,7, Daniel C Alexander 8, Michael O
Dada9, Iris Asllani10,11, Maruf Adewole12, and Farouk Dako13
1 McGill University, 3801 Rue University, Montreal, Quebec, H3A 2B4, Canada
2 Lawson Health Research Institute, 268 Grosvenor St, London, Ontario, N6A 4V2, Canada
3 Western University, 1151 Richmond Street, London, Ontario, N6A 3K7, Canada
4 Crestview Radiology Ltd., Victoria Island, Lagos, Nigeria
5 Mbarara University, Mbarara, Uganda
6 Flywheel Exchange, 1015 Glenwood Ave, Minneapolis, MN, 55405, USA
7 Washington University St Louis, MO, USA
8 University College London, Grower St, WC1E 6BT, UK
9 Federal University of Technology Minna, Niger State, Nigeria
10 University of Sussex, Brighton, BN1 9RH United Kingdom
11 Rochester Institute of Technology, 1 Lomb Memorial Dr, Rochester, New York, 14623, USA
12 University of Lagos, Lagos State, Nigeria
14 University of Pennsylvania, 3400 Spruce St, Philadelphia, 19104, USA
Udunna.anazodo@mcgill.ca
Abstract. Artificial intelligence (AI) can advance equitable access to high value
care and improve health outcomes across all settings. Since the burden of non-
communicable diseases is rising sharply in Sub-Saharan Africa (SSA) while de-
creasing in the Global North, the potential of AI is higher in such low resource
settings with greater disease burden and scarce healthcare availability. Yet, even
as global efforts are creating promising AI methods for medical imaging appli-
cations, it is unclear whether they can be deployed for clinical use in SSA due to
limitations in image quality and availability of curated medical data. This white
paper outlines an approach to address the challenges that reinforce disparities AI
applications in healthcare, focusing on imaging challenges unique to SSA. Over-
coming these challenges will enable full participation of SSA and other low re-
source settings in AI innovations thus making healthcare affordable and able to
address health outcomes across all settings.
Keywords: Artificial Intelligence, Brain MRI, Accessibility, Generalizability,
Data Governance, Fairness, Open Neuroimaging.
2
1 Background
1.1 The Potential for AI in Advancing Healthcare in Low Resourced Settings
Artificial intelligence (AI) can advance equitable access to high value healthcare and
improve health outcomes across all healthcare settings. Its potential is considered to be
higher in low resourced settings where the burden of disease is greater and the availa-
bility of healthcare infrastructure and highly skilled personnel is scarce [1]. Across low
resourced settings, particularly in low-and-middle income countries (LMICs), the inci-
dence, prevalence, morbidity, and mortality of non-communicable diseases (NCDs) are
growing due to the rapid economic (i.e., increase urbanization), demographic (i.e., ag-
ing population), and epidemiological (e.g., air pollution and changing behavioral trends
such as excess alcohol intake, sedentary lifestyle, and unhealthy diets) transitions [2].
In contrast to the Global North, where advancements in imaging technology have
contributed to a steady decrease in the burden of NCDs – such as cancer, stroke, heart
disease, dementia, etc., – the incidence rate of these disease in Sub-Saharan Africa
(SSA) is rising sharply. From 1990 to 2017, the total number of disability-adjusted life
years (DALYs) due to NCDs for all ages increased by 67% in SSA while the DALYs
due to NCDs decreased by 10.8% in the Global North [2], [3]. Even for glioblastoma,
the deadliest form of cancer and the most challenging to diagnose and treat, years of
extensive research to improve diagnosis, characterization, and treatment have de-
creased mortality rates in the Global North by 10-30% over the past 30 years [4]. These
research innovations have not translated to improvements in survival for individuals in
LMICs, particularly in SSA, where death rates rose on average by ~25% over the past
three decades [4]. These health disparities could be due to several overlapping factors;
delayed presentation [5], [6], high incidence of infectious disease comorbidities such
as HIV [7] (so called “double disease burden” [2]), severe shortage of healthcare infra-
structure including therapeutic options [5], [6], and lack of skilled expertise in diagnosis
(e.g., neuroradiologists and neuropathologists) and treatment (neurosurgeon, neuroon-
cologists, and medical physicists) [8], [9]. Despite these health system challenges, AI
has the potential to provide rapid diagnosis, speedup workflows, improve diagnostic
image quality and enable imaging-based country-specific and population-level assess-
ments of disease trends to close health outcome disparities [1].
1.2 Barriers to Fairness in AI Applications in Medical Imaging in Africa
We recently outlined the infrastructural, human capacity, and policy challenges that
hinder the robust AI ecosystem needed to empower low resourced settings to drive AI
solutions for the health of their population and overall society [1] (Fig. 1). Our recent
need assessment studies [9], [10] showed these challenges are intricately connected and
further exacerbated in medical imaging due to the higher cost of data acquisition (i.e.,
expensive scanners, multidisciplinary skilled personnel, large breadth and veracity of
data), storage, among other unique diagnostic imaging features (e.g., three-dimensional
data, embed personal identifiers). These barriers make AI by and large unattainable for
imaging applications in LMICs particularly in SSA, where the lack of political will to
3
invest in healthcare, especially in NCDs, along with the endemic unreliable network
connectivity, further compound these fundamental challenges.
Fig. 1. AI Challenges in healthcare in LMICs cut across infrastructural, capacity, and policy gaps.
Although global efforts to create state-of-the art AI methods for medical imaging
applications including several MICCAI grand challenges are promising for detection,
characterization, and treatment of NCDs, even the hard-to-diagnose-and-treat glioblas-
toma (i.e., BraTS [11]), it is unclear whether these AI methods can be implemented for
clinical use in SSA, given the wide use of lower quality imaging technology [1] , [9]
and more importantly, the limited availability of curated data (Fig. 2) [12].
Fig. 2. Illustration of the differences in quality and breath of imaging between Global North and
LMICs. (Top) representative 3T T1-weighted (T1w), T2-weighted (T2w), and enhanced T1-
weighted MRI scans from the SISS and BraTS collected largely from Europe and North America,
demonstrating the highly quality and plenty shot (data size and multispectral) and (Bottom) a
representative 0.3T T1-weighted (T1w) and T2-weighted (T2w) from Nigeria, demonstrating the
low quality and few shot nature of the dataset. (adapted from [12])
We recently demonstrated the need to expand publicly available medical imaging data
to include the lower quality and fewer sets of data from LMICs (Fig. 3) [12]. Using T1-
and T2-weighted Brain MRI data simulated to mimic the low resolution (1.2 x 1.2 x 6
mm) and nosier images more prevalent in SSA, we developed a robust model (SIGN,
4
Fig. 3) that considered these data challenges to effectively identify and segment stroke
lesions on the two sets of lower quality data, as well on conventional high resolution
isotropic (1.0 mm3) images typical in richer clinical settings [12].
In this white paper, we outline an approach to address key infrastructural, personnel
and policy challenges that reinforce disparities in application of AI in healthcare in
LMICs, focusing on imaging needs and challenges unique to SSA. If solved, this ap-
proach will enable the inclusion and full participation of SSA and other LMICs in AI
innovations to make healthcare affordable and drive the use of AI in addressing health
outcomes across all settings.
Fig. 3. A SimIlarity-weiGhed self-eNsembling (SIGN) framework (ours) outperformed existing
few-shot methods for segmentation of stroke lesions on low-quality few-shot MRI (T1-weighted
and T2-weighted) MRI (adapted from [12]).
2 Approach
2.1 The African Neuroimaging Archive (AfNiA)
We are developing AfNiA, a publicly available medical imaging archive that will
provide a comprehensive AI ecosystem (Fig. 4) to address the key barriers described
above and equip SSA communities with the tools to participate in global efforts to de-
velop and rapidly deploy AI imaging innovations, from model-to-clinic.
Fig. 4. The AfNiA approach outlining its 12 nodes for democratizing AI in healthcare.
5
Starting with the data, AfNiA will aggregate existing brain MRI scans and relevant
clinical attributes from multiple clinics across Africa into an enterprise data manage-
ment platform, Flywheel (https://flywheel.io/). AfNiA will leverage the comprehensive
data management solutions in Flywheel to develop (or adapt pre-existing) tools for data
curation, quality control, image processing and image labelling/annotation [13], [14]
(Fig 4) to accommodate large-scale MRI data from various scanners (vendor, model,
0.3 to 3 T field strength) at ~22 institutions in Nigeria, Ghana, and Ethiopia. This will
provide a rich repository of imaging data carefully curated to preserve the heterogeneity
of real-world clinical data and the inherent complexity of clinical presentations of neu-
rological disorders including overlapping pathological features and neurological-nor-
mal (i.e., patient scans deemed normal) scans. The containerized algorithms for image
processing, annotation, and image analysis developed with Flywheel or the broader re-
search community for AfNiA will be open-source and sharable through public reposi-
tories (i.e., XNAT or GitHub) to meet open science recommendations.
AfNiA is being developed in two phases. Phase 1 will establish the framework (Fig.
4) in collaboration with regional partners using ~1000 locally sourced multi-center
brain MRI images from around Lagos, Nigeria to solve lingering infrastructure (includ-
ing data quality via image quality enhancement options[15]), personnel, and policy
challenges (Figs 1 & 4). Phase 2 will expand the utility of AfNiA for generalizability
efforts to include aggregation of >5000 brain MRI images from other regions in Nigeria
and SSA starting with Ghana and Ethiopia, focusing on addressing infrastructure con-
straints and local policy challenges. To enable participation from centers without elec-
tronic archiving and storage (PACS) or internet connectivity, which are often the least-
resourced and more remote settings in the region [1], AfNiA will establish end-to-end
encrypted data transfer via registered mail delivery of secure portable devices from
clinics to a regional central data center (Fig. 4). An approach for data ingestion and
preprocessing compliant with FAIR principles[16] and conventional data privacy and
protection regulations will be developed to streamline the series of time-consuming
data curation steps, from anonymization to data quality verification and crowdsourced
data annotation [1] (Fig. 4).
To address the lack of region-specific data governance provisions, policies for ethi-
cal use are being developed to govern access to curated data and ensure that AI solu-
tions created using AfNiA will directly benefit the local communities who own and
provide the data [1]. This effort is crucial in LMICs, where medical data is overwhelm-
ingly provided from out-of-pocket patient expenses, and as such require direct benefit
considerations. Because of the multidisciplinary nature of medical imaging and the lim-
ited and often isolated expertise in the region, AfNiA will leverage its data governance
mechanism and the growing imaging network in Africa via The Consortium for Ad-
vancement of MRI Research and Education (CAMERA) [9] to link users who develop
AI tools to local clinics that supply data to enable co-creation of AI solutions and es-
tablish a global peer-to-peer collaborative network (Fig. 4), capable of turning African
colleagues from providers to expert developers.
To evaluate the utility of AfNiA and all its nodes (Fig. 4), we will release up to 300
annotated glioma datasets from AfNiA to support global efforts to improve brain tumor
6
diagnosis and treatment. In partnership with the MICCAI Brain Tumor Segmentation
(BraTS) Challenge, we will enrich the BraTS challenge data and enable evaluation of
proposed state-of-the-art models for tumor segmentation [1]. Over the next three years,
we plan to release series of annotated neurological data (i.e., pituitary adenoma, is-
chemic stroke, epilepsy, etc.), expand beyond brain MRI to other modalities and organs,
and populate the platform with diverse data (pathology, laboratory, demographics, etc.)
for future multi-omics and epidemiology studies.
To meaningfully address gaps in local AI imaging capacity from data collection to
AI model development, evaluation, and clinical translation [1] (Fig 4) we are integrat-
ing the SPARK (Sprint AI Training for African Medical Imaging Knowledge Transla-
tion) program in AfNiA, as a train-the-trainer workshop and hackathon (Fig.5). SPARK
is aimed at upskilling a new generation of African AI experts who will through collec-
tive intelligence (i.e., medical physics + computer science + radiology + neurology/neu-
rosurgery) rapidly turn knowledge-to-action, for example, by fully participating in up-
coming imaging grand challenges. The first SPARK program is planned to coincide
with release of the AfNiA brain tumor data in 2023, ahead of the BraTS 2023 challenge.
Fig. 5. The SPARK training model to rapidly transform regional AI imaging capacity.
In general, Our proposed approach implements current recommendations for, 1) med-
ical imaging data preparation for AI[17]–[20], 2) development and evaluation of AI-
based imaging diagnostics[21], and 3) fair ethical use and sharing of clinical imaging
data for AI[22], [23] to effectively provide a sustainable solution for democratizing AI
in medical imaging.
3 Conclusion
We propose a unified strategy to close gaps in AI in healthcare infrastructure, capac-
ity, and policy in low resourced settings. By focusing on AI medical imaging needs and
challenges in SSA, we will build a robust solution that can be adapted across LMICs,
as well as to non-imaging AI health needs. In lieu of limited infrastructure and capacity
for federated data systems, our centralized repository provides a foundational base for
inclusion of LMICs in federated learning efforts (‘quasi-data federation’) towards fu-
ture data federation as information technology capacity becomes more available in the
region.
7
References
[1] U. C. Anazodo, M. Adewole, and F. Dako, “AI for Population and Global
Health in Radiology,” Radiol. Artif. Intell., p. e220107, Jun. 2022.
[2] R. M. Gyasi and D. R. Phillips, “Aging and the Rising Burden of
Noncommunicable Diseases in Sub-Saharan Africa and other Low- and
Middle-Income Countries: A Call for Holistic Action.,” Gerontologist, vol. 60,
no. 5, pp. 806–811, Jul. 2020.
[3] H. N. Gouda et al., “Burden of non-communicable diseases in sub-Saharan
Africa, 1990–2017: results from the Global Burden of Disease Study
2017,” Lancet Glob. Heal., vol. 7, no. 10, pp. e1375–e1387, Oct. 2019.
[4] GBD 2016 Brain and Other CNS Cancer Collaborators, “Global, regional, and
national burden of brain and other CNS cancer, 1990-2016: a systematic
analysis for the Global Burden of Disease Study 2016.,” Lancet. Neurol., vol.
18, no. 4, pp. 376–393, Apr. 2019.
[5] C. A. Ndubuisi, W. C. Mezue, M. Nzegwu, O. Okwunodulu, G. Ejembi, and S.
C. Ohaegbulam, “The Challenges of Management of High-grade Gliomas in
Nigeria,” J. Neurosci. Rural Pract., vol. 8, no. 3, pp. 407–411, 2017.
[6] R. O. Akinyemi et al., “Stroke in Africa: profile, progress, prospects and
priorities,” Nat. Rev. Neurol., vol. 17, no. 10, pp. 634–656, 2021.
[7] A. Kelly, P. Lekgwara, and S. Mda, “The epidemiology and outcome of
patients admitted for elective brain tumour surgery at a single neurosurgical
centre in South Africa,” Interdiscip. Neurosurg., vol. 21, p. 100750, 2020.
[8] S. Mukhopadhyay et al., “The global neurosurgical workforce: a mixed-
methods assessment of density and growth,” J. Neurosurg., vol. 130, no. 4, pp.
1142–1148, 2019.
[9] U. C. Anazodo et al., “A Framework for Advancing Sustainable MRI Access
in Africa,” medRxiv, vol. 18, p. 2022.05.02.22274588, May 2022.
[10] A. Elahi et al., “Overcoming Challenges for Successful PACS Installation in
Low-Resource Regions: Our Experience in Nigeria,” J. Digit. Imaging, vol. 33,
no. 4, pp. 996–1001, 2020.
[11] B. H. Menze et al., “The Multimodal Brain Tumor Image Segmentation
Benchmark (BRATS),” IEEE Trans. Med. Imaging, vol. 34, no. 10, pp. 1993–
2024, Oct. 2015.
[12] D. Zhang, R. Confidence, and U. Anazodo, “Stroke lesion segmentation from
low-quality and framework,” in MICCAI 2022, pp. 1–10.
[13] F. Kofler et al., “BraTS Toolkit: Translating BraTS Brain Tumor Segmentation
Algorithms Into Clinical and Scientific Practice,” Front. Neurosci., vol. 14, p.
125, Apr. 2020.
[14] C. Davatzikos et al., “Cancer imaging phenomics toolkit: quantitative imaging
analytics for precision diagnostics and predictive modeling of clinical
outcome.,” J. Med. imaging, vol. 5, no. 1, p. 11018, Jan. 2018.
[15] M. L. de Leeuw den Bouter, G. Ippolito, T. P. A. O’Reilly, R. F. Remis, M. B.
van Gijzen, and A. G. Webb, “Deep learning-based single image super-
resolution for low-field MR brain images,” Sci. Rep., vol. 12, no. 1, pp. 1–10,
8
2022.
[16] M. D. Wilkinson et al., “The FAIR Guiding Principles for scientific data
management and stewardship.,” Sci. data, vol. 3, p. 160018, Mar. 2016.
[17] O. Diaz et al., “Data preparation for artificial intelligence in medical imaging:
A comprehensive guide to open-access platforms and tools,” Phys. Medica, vol.
83, no. November 2020, pp. 25–37, 2021.
[18] J. Li et al., “A Systematic Collection of Medical Image Datasets for Deep
Learning,” 2021.
[19] K. Clark et al., “The Cancer Imaging Archive (TCIA): Maintaining and
Operating a Public Information Repository,” J. Digit. Imaging, vol. 26, no. 6,
pp. 1045–1057, 2013.
[20] M. J. Willemink, W. A. Koszek, M. S. C. Hardell, M. S. J. Wu, D. L. Rubin,
and M. S. M. P, “Preparing Medical Imaging Data for Machine Learning,”
Radiology, vol. 295, pp. 4–15, 2020.
[21] D. B. Larson, H. Hugh, D. L. Rubin, N. Irani, J. R. Tse, and C. P. Langlotz,
“Regulatory Frameworks for Development and Evaluation of Artificial
Intelligence–Based Diagnostic Imaging Algorithms: Summary and
Recommendations,” J. Am. Coll. or Radiol., vol. 18, pp. 413–424, 2021.
[22] D. B. Larson, D. C. Magnus, M. P. Lungren, N. H. Shah, and C. P. Langlotz,
“Ethics of using and sharing clinical imaging data for artificial intelligence: A
proposed framework,” Radiology, vol. 295, no. 3, pp. 675–682, 2020.
[23] J. Wawira Gichoya, L. G. McCoy, L. A. Celi, and M. Ghassemi, “Equity in
essence: A call for operationalising fairness in machine learning for
healthcare,” BMJ Heal. Care Informatics, vol. 28, no. 1, pp. 2020–2022, 2021.
