AfaraBook a call
← All articles

November 18, 2026

Shahin Sayed on What It Takes to Validate Cancer AI in Kenya

Shahin Sayed on What It Takes to Validate Cancer AI in Kenya

Abstract digital pathology tiles moving through an expert review gate, in forest green, teal, sand and muted amber

Meta description: How clinical data annotation shaped a Kenyan colorectal cancer AI study, and what its results do, and do not, show about model evaluation.

Clinical data annotation is the expert-labelled ground truth a model is trained and judged against. A Kenyan colorectal cancer study shows how much a model’s credibility depends on who creates those labels and how they are checked. In this case, a computational tool prioritised tissue-image tiles for review, but a pathologist made the final decision on each label.

Who Sayed is and why pathology is the bottleneck

Professor Shahin Sayed is Professor of Pathology at Aga Khan University Hospital, Nairobi; Department of Pathology, Aga Khan University, Nairobi. Sayed’s work includes training and strengthening the capacity of pathologists and researchers across sub-Saharan Africa.

The University of Michigan Center for Global Health Equity profile describes Sayed’s work on Kenya’s first clinically annotated cancer biorepository. The resource is intended to support translational research, biomarker validation and AI-enabled computational pathology. Sayed’s research also includes validating AI and machine-learning approaches for colorectal cancer diagnosis and prognostication in low-resource settings, and developing methods to address pre-analytic variation and image-domain artifacts in digital pathology.

The workforce context is stark. The study reports that many countries in sub-Saharan Africa have fewer than one pathologist per million people, compared with roughly one pathologist per 20,600 people in the United States. The authors say that at current workforce growth rates it would take around 400 years for Africa to reach the high-income-country ratio. These figures are from the paper, not a forecast made by this article.

This shortage makes tools that help pathologists focus their attention worth investigating. It does not make the need for expert pathology or strong safeguards any less important.

What the study set out to test

In a paper co-senior-authored by Professor Sayed and Mansoor N. Saleh, the team asked whether a colorectal cancer screening pipeline first developed in a North American cohort could be adapted to Kenyan tissue and practice settings.

The study, published in the Journal of Pathology Informatics on 20 July 2026, was a retrospective, multi-institutional evaluation. It included two deliberately contrasting sites: Aga Khan University Hospital Nairobi, a tertiary teaching and referral centre, and Tenwek Hospital, a non-academic community hospital in rural Bomet.

The source cohort included 136 histologically confirmed colonic adenocarcinoma cases diagnosed between 2016 and 2019, plus 20 normal controls. After a gastrointestinal-trained pathologist reviewed slide quality, 95 cancer cases and 19 controls remained. Slides were digitised at 40x using a Grundium Ocus 40 scanner and stored as SVS whole-slide images.

The case-level reference labels came from the original clinical pathology diagnoses, made by board-certified pathologists before digitisation. That separation matters: the reference diagnosis was not produced by the model being evaluated.

The annotation step most people skip

The study’s central lesson for clinical data annotation is that a machine-assisted workflow still needs expert decisions at its foundation.

First, VIPR/SIVQ, a computational feature-enrichment approach developed and validated in a North American cohort, prioritised candidate 512 × 512-pixel tiles likely to contain adenocarcinoma-associated morphology. Then a gastrointestinal-trained pathologist reviewed the candidate tiles and placed them into adenocarcinoma-containing or benign/non-neoplastic libraries.

Across 70 of the 95 quality-controlled cancer cases, roughly 200 to 500 tiles were labelled per case. The key distinction is easy to miss: the tool helped find candidate regions; it did not decide whether a tile was cancerous. Final tile labels, and therefore the tile-level ground truth, were determined entirely by the pathologist.

Abstract tiles selected by a computational filter and then sorted through a human expert review gate

This order matters. Automated prioritisation can make review more focused, but it is not the same as clinical judgement. And the case-level ground truth had a separate source: the original diagnoses made before the algorithm was involved.

The team used the curated tiles to retrain a compact MobileNet-based convolutional neural network in TensorFlow/Keras. The model was exported as a single portable file intended to run on commodity hardware without a dedicated GPU. To produce a case-level score, the workflow ranked the tiles by malignancy probability, averaged the top 5 percent and classified the case as malignant at a threshold of 0.5.

What the numbers showed, and what they do not

The paper reports results from 14,452 tiles drawn from 70 adenocarcinoma cases and 19 benign controls. At tile level, the reported sensitivity was 0.9548, specificity 0.9926, precision 0.9828, accuracy 0.9810, negative predictive value 0.9803, F1 score 0.9686 and AUROC 0.9662.

At case level, Top-K mean pooling produced an AUROC of 1.0. The model identified all adenocarcinoma cases in the evaluation and correctly classified all 19 controls.

Those results deserve careful reading. The authors state that the 19 control cases also informed operating-point selection: the team selected a threshold to maintain 100 percent case-level sensitivity while minimising false positives among those controls. So the reported specificity describes the achieved operating point, not an independent estimate of specificity. The perfect case-level separation should also be interpreted cautiously.

That distinction is important for medical model evaluation. A strong result on a defined task can support further study, but it does not establish performance across other hospitals, scanners, patient groups or diagnostic questions. The study does not show that the pipeline is clinically deployed or approved.

The limits the authors wrote down themselves

The paper identifies several limits that bound what can be concluded:

  • A modest cohort: there were 95 quality-controlled cases and a small holdout set, with only 19 negative controls available for false-positive optimisation. The study had no formal power calculation or pre-specified non-inferiority margin.
  • A narrow task: the model classified adenocarcinoma versus benign tissue. The evaluation did not systematically test difficult diagnostic mimics, grading, subtyping, mucinous variants, dysplasia or inflammatory backgrounds.
  • Local retraining: the pipeline was retrained on Kenyan data and evaluated on an internal holdout. The authors describe this as local adaptation with internal validation, not an evaluation of a frozen model on fully independent data.
  • Uncertain generalisability: results came from two Kenyan institutions and one scanner platform. Performance beyond those settings is unknown.

In practical terms, this is not evidence that a model trained elsewhere will work unchanged. Nor is it evidence that this adapted version will perform the same way on every slide or diagnostic challenge. It is a report of a specific adaptation and evaluation, with limits the authors make explicit.

Four governance rules for deployment

The authors propose four pillars to guide any future deployment:

  1. Pathologist final sign-out. The model may support screening or triage; the pathologist retains final diagnostic authority.
  2. Transparency. Teams should state what the tool is intended to do and where its limitations lie.
  3. Traceability and auditability. Algorithm outputs should be documented so teams can review what the system produced and how it was used.
  4. Ongoing performance monitoring. Teams should watch for drift and changes in performance over time.

The paper also argues against a double standard: deployment in a low- or middle-income country should not mean accepting weaker safeguards or lower performance thresholds than would be accepted in a high-income setting. The authors align this position with the WHO’s ethical principles for AI in health.

Data governance was part of the study design, too. The authors report that images were hosted on a cloud server with redundant storage in a South African data centre, consistent with Kenya Data Protection Act requirements. The study received ethics approvals from Aga Khan University (2023/ISERC-118), Tenwek Hospital (2023-0022) and the University of Michigan (HUM00240708), plus a research licence from Kenya’s National Commission for Science, Technology and Innovation (NACOSTI/P/24/32424). Funding was provided under NIH award U01CA287852.

What Afara takes from this

For Afara, the study reinforces the importance of credentialed clinicians from the global south producing the ground truth that medical AI depends on. Here, the pathologist kept final authority over labels while the model handled tile scoring and case-level triage.

For teams working on healthcare data labeling or medical AI evaluation, the full paper is a useful reference for considering how expert review, local context and limits should be documented.

FAQ

Why are there so few pathologists in Africa?

The study reports that many sub-Saharan African countries have fewer than one pathologist per million people. It contrasts this with roughly one pathologist per 20,600 people in the United States and says that current workforce growth rates would take around 400 years for Africa to reach the high-income-country ratio.

Can AI make cancer screening faster in low-resource settings?

AI may help prioritise cases or regions for review, but this study evaluated a locally retrained colorectal cancer pipeline; it did not establish clinical deployment. Its authors emphasise pathologist sign-out, transparency, auditability and continuous monitoring as requirements for future deployment.

What makes a good ground truth dataset for medical AI?

In this study, a gastrointestinal-trained pathologist determined tile labels, while case-level labels came from original clinical pathology diagnoses made before digitisation and algorithm development. Clear separation between model outputs and reference labels helps make evaluation more credible.

For further detail, read the full study and Professor Sayed’s profiles at Aga Khan University and the University of Michigan Center for Global Health Equity.