The resource-constrained health systems of many low- and middle-income countries (LMICs) offer many medical imaging use cases where machine learning, properly deployed, can greatly improve quality of care and health outcomes for millions of underserved individuals. Global Health Labs, Inc. (GH Labs) is honored to host a MICCAI tutorial focused on effective use of machine learning (ML) for medical use-cases in LMICs. We invite the MICCAI community to join us in spotlighting this important, interesting, and high value set of challenges for medical imaging and machine learning.
September 18, 2022
8:00 – 11:30 AM (SGT time)
MICCAI, Singapore
Please visit the MICCAI Conference website to register and obtain streaming instructions for virtual attendance
This tutorial seeks to articulate the unique opportunities, constraints, and concerns that underlie efforts to develop and deploy ML-based health care solutions in low- and middle- income countries (LMICs). The challenge is to unite three disparate realms: ML, Medical, and the system limitations of LMICs.
The intersection of these three disparate realms defines a unique space, whose overlapping constraints and features must be understood and addressed if a deployment is to succeed.
In this tutorial, we will offer actionable insights to help define, develop, and deploy effective ML solutions in this intersectional ML-Medical-LMIC space. Our perspective is that of ML practitioners specialized in LMIC health care delivery. We will articulate various general issues, illustrating each one with concrete examples drawn from actual projects including malaria diagnosis, lung ultrasound, and cervical cancer screening.
Topics will include: the central role of field partners and subject matter experts; tailoring ML metrics tailored to the use case; issues of annotating and cleaning data; and a project case study (cervical cancer screening).
Participants will leave with insights about core issues in the ML-Medical-LMIC space, and concrete methods for how to address them to develop effective solutions.
This half-day tutorial will include several talks, and ample time for Q & A as well as networking. Some topics are given below.
Welcome and introduction of themes
Keynote
Understanding the use case—why and how to connect with domain expertise
Key issues in data and annotation
Break
RISE (recruiting LMIC researchers)
Use case-focused metrics and model design
Cervical cancer screening with ML—a case study
Closing remarks
Field use cases combine medical demands (e.g., performance specs) with environmental demands (e.g., low cost, infrequent patient contacts, tough field conditions, and difficult and expensive data collection). Collaborations driven by field partners, who know these field constraints (but who don't know ML) are key to meeting complex product requirements. ML practitioners may best be viewed as akin to surgeons, called in to apply specific expertise in a context defined by other experts.
ML practitioners tend to use certain standard metrics to guide and assess algorithm development. This is due to institutional habits and to incentives (e.g., to be accepted, a paper often needs to include a SOTA bake-off with standard datasets). Often, however, algorithm development is best served by defining metrics that concretely capture the requirements of the use case. We discuss the hows and whys of such tailored metrics, with examples drawn from automated malaria diagnosis and other projects.
Unlike pre-cleaned and standard datasets often used in ML research, datasets from the LMICs are rarely (never) ready to drop into ML frameworks. Success of a project is often determined by the data preparation. For example, the annotations must be tailored to match the particular use case and algorithm target. We highlight key issues of annotation and cleaning, with examples from public health predictive modeling, vitamin A testing, and other projects.
Ultrasound promises to be a powerful, enabling modality for many LRS use cases, and illustrates multiple lessons relevant to our theme. We’ll discuss ultrasound work as a case-study to highlight various issues in the ML-Medical-LMIC space. Examples include complexities of annotations, conflicting high-income country and LMIC use cases, and effects of use case on training set tuning.
Field use cases combine medical demands (e.g., performance specs) with environmental demands (e.g., low cost, infrequent patient contacts, tough field conditions, and difficult and expensive data collection). Collaborations driven by field partners, who know these field constraints (but who don't know ML) are key to meeting complex product requirements. ML practitioners may best be viewed as akin to surgeons, called in to apply specific expertise in a context defined by other experts.
ML practitioners tend to use certain standard metrics to guide and assess algorithm development. This is due to institutional habits and to incentives (e.g., to be accepted, a paper often needs to include a SOTA bake-off with standard datasets). Often, however, algorithm development is best served by defining metrics that concretely capture the requirements of the use case. We discuss the hows and whys of such tailored metrics, with examples drawn from automated malaria diagnosis and other projects.
Unlike pre-cleaned and standard datasets often used in ML research, datasets from the LMICs are rarely (never) ready to drop into ML frameworks. Success of a project is often determined by the data preparation. For example, the annotations must be tailored to match the particular use case and algorithm target. We highlight key issues of annotation and cleaning, with examples from public health predictive modeling, vitamin A testing, and other projects.
Ultrasound promises to be a powerful, enabling modality for many LRS use cases, and illustrates multiple lessons relevant to our theme. We’ll discuss ultrasound work as a case-study to highlight various issues in the ML-Medical-LMIC space. Examples include complexities of annotations, conflicting high-income country and LMIC use cases, and effects of use case on training set tuning.
Welcome and introduction of themes
Keynote
Understanding the use case—why and how to connect with domain expertise
Key issues in data and annotation
Break
RISE (recruiting LMIC researchers)
Use case-focused metrics and model design
Cervical cancer screening with ML—a case study
Closing remarks
Groesbeck Parham is professor of gynecologic oncology at the University of North Carolina. He founded the Cervical Cancer Prevention Program in Zambia, where he developed a system for cervical cancer screening which has been scaled across the country and was recently leveraged to create a diagnostic based on machine learning. He serves as the ‘Clinical Expert for the Cervical Cancer Elimination Initiative' at WHO.
Noni Gachuhi has over 21 years of experience in public health program delivery in HIV/AIDS prevention, reproductive health/family planning and malaria programming. She has lived and worked in Kenya, Zimbabwe, Uganda, Rwanda, and India, designing, implementing, and monitoring public health programs.
Matt Horning leads the machine learning team at GH Labs. He has worked on technologies to address global health needs, primarily in optics and computer-vision based diagnostic techniques, since 2010. Applications have included malaria diagnosis, cervical cancer screening, and lung ultrasound.
Andrea Lara is the director and founder of BiomedLab, Galileo University, Guatemala's first biomedical engineering research laboratory. Her research field is medical image processing and analysis, focused mainly on cardiac CT. She is a Ph.D. candidate at the Graz University of Technology, has a master's degree in biomedical engineering from the University of Lübeck, and worked as a research assistant at the Fraunhofer Institute for Biomedical Engineering.
Courosh Mehanian received a physics PhD from Cornell University. He has worked for three decades in artificial intelligence, having held positions at Boston University, MIT Lincoln Laboratory, GH Labs, and now the University of Oregon, where his lab applies ML and computer vision to automated medical image understanding.
Charles Delahunt has 9 years of experience applying ML to global health projects including malaria diagnosis, vitamin A testing via eye videos, ultrasound, helminth egg detection, and pregnancy risk assessment. He has also held a postdoc in the University of Washington’s applied math department, focused on ML methods.
Global Health Labs, Inc. (GH Labs) innovates to reduce health disparities, especially in low- and middle-income countries. As a nonprofit corporation fully funded by Gates Ventures (the private office of Bill Gates), we closely partner with the Bill & Melinda Gates Foundation and leaders across sectors to develop technology solutions that address unmet needs in:
Recent innovations from GH Labs resulted in several commercially available rapid COVID-19 antigen tests, an AI-enabled smartphone app to detect cervical cancer, and a portable cold chain device that enables vaccine delivery to remote communities—among many other tools and technologies advanced for the people who need them most.
(+ 1) 425.777 .9687
Global Health Labs
14360 SE Eastgate Way
Bellevue, WA, 98007-6462
Presentation Slides
MICCAI for Global Health Tutorial 2022
Journal Article
Journal Article
Computer-automated Malaria Diagnosis and Quantitation
Journal Article
Fully-automated patient-level Malaria assessment
Journal Article
Performance of a fully-automated system
Journal Article
