Partnership powers AI-driven cancer diagnosis in Kenya

When Mansoor Saleh, an oncologist, and Shahin Sayed, a pathologist, encountered a critical barrier to cancer care at Aga Khan University in Nairobi, Kenya, they already had the expertise and resolve to solve it. What they needed was a like-minded partner who could help them unlock cutting-edge resources and transform their ideas into reality. 

Enter Akbar Waljee, MD, and this team at the Center for Global Health Equity at the University of Michigan —not as leaders directing the work, but as enablers connecting researchers across continents to tackle one of global health's most pressing challenges. 

The problem: A growing cancer crisis 

Colorectal cancer ranks as the third most commonly diagnosed cancer worldwide and the second leading cause of cancer deaths. While these statistics are sobering everywhere, the situation in Africa presents unique challenges. Cancer rates across the continent are rising steadily, as is the growing incidence of colon cancer in patients below 40 years, yet the specialized pathology services essential for diagnosing and treating these cancers remain scarce. 

The gap is stark. Accurate cancer diagnosis requires trained pathologists who can examine tissue samples under microscopes and identify specific cellular features that indicate cancer type, stage, and aggressiveness – attributes needed to select the most appropriate treatment approach. But the presence of cancer cells in samples is not always immediately apparent — in ambiguous cases, pathologists must spend considerably more time reviewing slides and using special stains to reach a confident diagnosis. In many countries, however, pathologists are few and far between. In Kenya, for example, some estimates count as few as 0.13 pathologists per 100,000 people — or just one pathologist for every 1 million people —making it one of the lowest ratios globally. The specialized training to become a pathologist takes years, and the infrastructure and equipment needed for modern pathology practices can be prohibitively expensive for hospitals in low- and middle-income settings. 

The innovation: Teaching machines to see cancer 

Ulysses Balis, a University of Michigan professor of pathology and member of the center, is a leading expert in artificial intelligence and machine learning applications in pathology. As the primary architect of the SIVQ/VIPR (Spatially Invariant Vector Quantization/Vectorizing Pattern Recognition) computational pipeline — first developed in 2012 and continuously refined since — Balis created a sophisticated image processing pipeline for pathology images that holds the potential to fundamentally transform how AI can facilitate and expedite the analysis of tissue biopsies sent for histopathologic analysis. 

Working alongside researchers at Aga Khan University and University of Michigan colleagues, including Arvind Rao, Jerome Cheng, and Akbar Waljee, Balis's technology addresses one of pathology's most time-consuming challenges: diagnostic uncertainty. The VIPR system serves as an intelligent pre-processor that enables subsequent machine learning algorithms to assess cancer likelihood with remarkable precision, giving pathologists a powerful pre-screening tool that facilitates critical decision support, by assisting in the prioritization of cases in need of further review and consultation. In total, adding this computational workflow can accelerate diagnoses, leading to expedited and better targeted treatment. 

The technology works by examining the same hematoxylin and eosin-stained slides that pathologists have used for more than a century, but it does so at the pixel level in a fundamentally different way. Rather than simply analyzing colors, VIPR biologically contextualizes every single pixel in an image. The system examines each pixel and its surrounding structure, then assigns biological meaning based on pre-trained classifiers: this pixel is part of a nucleus, this one is part of an epithelial cell, this one is part of a vascular structure, this one belongs to a malignant cell. 

This contextualization converts each pixel from a simple three-dimensional color value (red, green, blue) into a high-dimensional biological space of 10, 20, 30, or even 40 dimensions. This enriched, spatially-mapped data then feeds into conventional machine learning algorithms, which can converge much more efficiently on accurate diagnoses because they're starting with biological context rather than raw color information. 

What makes VIPR particularly suited for settings with limited resources is its semisupervised training approach. Unlike traditional convolutional neural networks that require tens of thousands of images to learn patterns, VIPR can be trained by pathologists simply clicking on relevant areas and identifying what they are. This "handcrafted features" approach means the system can sometimes converge on accurate, generalizable solutions with as few as four or five fields of view from a single case—a critical advantage when working with rare diseases or in settings where large datasets simply don't exist. 

Once the AI identifies distinct tissue regions through VIPR's contextualization, additional algorithms analyze these areas to predict not just whether cancer is present, but also how aggressive the disease might be and what treatment approaches could work best. 

The partnership approach: Amplifying local leadership 

This is where the Center for Global Health Equity's role becomes transformative. Rather than simply providing suggestions and long-distance advice, the center functions as what its leaders call a "partnership gateway" — connecting researchers who are already doing innovative work with the resources, expertise, and networks they need to scale their impact. 

Through a National Institutes of Health U01 grant funded in September 2023 and running through August 2026, the center brought together a team spanning two continents. The collaboration includes Aga Khan University in Nairobi, Kenya, the University of Michigan, and Tenwek Hospital — a faith-based community hospital in rural Bomet, Kenya, under the leadership of Robert Parker. This partnership unites oncologists, pathologists, surgeons, statisticians, and computer scientists, each bringing essential expertise to develop solutions specifically designed for the Kenyan context. 

The center's approach embodies what global health partnership should look like in 2025: grounded in mutual respect, focused on local leadership, and designed to build lasting capacity rather than temporary fixes. 

The research: Two critical aims 

The project pursues two interconnected goals, each building toward a comprehensive diagnostic tool. The project employs advanced computational methods to analyze tissue samples from nearly 140 colorectal cancer cases diagnosed between 2009 and 2021 at the Aga Khan University Hospital and Tenwek Hospital in Kenya. The AKU team identifies, prepares, and verifies hematoxylin and eosin-stained slides of all eligible samples, along with normal colon tissue controls. At Tenwek Hospital, the slides are anonymized and digitized using a Grundium digital microscope and scanner donated by the University of Michigan. The digitized images are then securely transmitted to the Michigan team for analysis. 

Aim 1 focuses on adapting and validating Dr. Balis's machine learning algorithm for colorectal cancer diagnosis using tissue samples from the local Kenyan population. At the University of Michigan, the digitized images are analyzed using the VIPR AI algorithms. By segmenting and classifying images at the pixel level, the system enhances diagnostic speed and accuracy, applying the SIVQ/VIPR pipeline to identify cancer features and ultimately providing a probability assessment: the likelihood of cancer being present or absent. 

Aim 2 takes the work further by developing methods to predict cancer prognosis — essentially, how the disease is likely to progress. Using artificial intelligence and machine learning technologies, the team identifies specific image features, such as cellular shape and structure, that correlate with aggressive disease. This information, combined with other clinical factors, helps doctors understand not just whether a patient has cancer, but how urgently they need treatment and what their likely outcomes might be. 

To date, over 150 H&E-stained colon cancer slides, as well as normal colon tissue controls, have been digitized and uploaded. Preliminary demonstrations of the platform's capability to identify cancer or the absence thereof have been very positive. 

Beyond research: Building infrastructure 

Perhaps the most exciting development emerging from this collaboration extends beyond the original grant. The path to implementing AI diagnostics requires first establishing digital pathology workflows—itself a major undertaking that involves not just sophisticated algorithms, but the practical realities of deploying new technology and transforming established practices. 

The challenges are significant, particularly in low- and middle-income country settings. Beyond the technical aspects of getting pathologists to shift from microscopes to computer screens, there are logistical and infrastructure hurdles to overcome. Establishing the systems for storing and analyzing digital images, ensuring adequate connectivity for transmitting large files, and navigating the practical requirements of bringing new technology into existing healthcare settings all require careful planning and sustained effort. 

The team's strategy has been deliberately sequential: establish digital pathology workflows first, then layer on AI capabilities once the infrastructure and human practices are firmly in place. Attempting to deploy both simultaneously would create too many challenges for a department to absorb effectively. By getting pathologists comfortable with digital workflows now—training them on the new systems, establishing routines, and solving logistical challenges—the groundwork will be laid for seamlessly introducing AI tools in the future. 

Through extensive meetings and collaborative planning, Aga Khan University has used lessons from the University of Michigan's own digital pathology implementation to build a compelling case for its institutional investment. AKU leadership has now committed to launching a comprehensive Digital Pathology Program that will modernize workflows, increase efficiency, and expand diagnostic capacity across their network. 

The momentum continues to build. Last year, Saleh presented the early findings at the U01 meeting in Mauritius, and recently, Sayed presented the project at the AORTIC meeting in Tunis, showcasing AKU's journey into digital workflow. The project was enthusiastically received, highlighting the broader regional interest in these approaches and positioning Aga Khan University as a leader in bringing advanced diagnostic technologies to Kenya and the wider East African region. 

This transformation exemplifies the "jetpack effect" that strategic partnerships can provide. The Center for Global Health Equity didn't just fund a research project — it catalyzed institutional change that will benefit thousands of patients for years to come. 

Scalable solutions for lasting impact

The implications reach far beyond colorectal cancer in Kenya. Because the system relies on open-source software and can be deployed through cloud-based platforms, it's designed for widespread adoption and adaptation. The scalable, cloud-based technologies make it particularly suited to low-resource settings. The same approach can be integrated into medical education programs, helping train the next generation of pathologists while providing them with powerful diagnostic support tools. 

Moreover, the methodology can extend to other cancers and noncommunicable conditions that disproportionately affect people in low- and middle-income countries. The research team is optimistic about the project's potential to be adapted to detect cancers such as breast, cervical, and prostate cancer, paving the way for tools that not only detect cancer but guide treatment decisions. Each application builds on the same fundamental approach: leveraging artificial intelligence to overcome barriers of specialized training and expensive technology, making it possible for a variety of cancers to be diagnosed accurately and treated appropriately, contributing to better patient outcomes. 

A model for global health collaboration

This colorectal cancer diagnostic project demonstrates what's possible when global health partnerships are built on the right foundation. The Center for Global Health Equity serves as a connector, an amplifier, and an enabler that helps researchers like those at Aga Khan University and Tenwek Hospital access University of Michigan's cutting-edge expertise in AI and pathology, along with implementation science resources. 

The partners in Kenya are the innovators. They understand their communities' needs, they've identified the problems that matter most, and they're developing solutions that will actually work in their contexts. What the center provides is a platform — a way to unlock resources, forge connections, and accelerate the path from promising ideas to life-saving reality. 

As cancer rates continue rising across Africa and other low-income regions, this model of partnership-driven innovation offers hope. By combining local expertise with global resources, grounding research in community needs and focusing on scalable, affordable solutions, collaborations like this one chart a path toward genuine health equity.

 The work happening in Kenya right now isn't just about diagnosing cancer more accurately. It's about reimagining how global health partnerships can function — not as one-way transfers of knowledge and resources, but as true collaborations where every partner brings essential expertise and all benefit from shared success to improve the lives of the communities we live in.

—— 

Learn more about the project: Leveraging artificial intelligence/machine learning-based technology to overcome specialized training and technology barriers for the diagnosis and prognostication of colorectal cancer in Africa