Lessons learned from the validation of a machine learning-based colorectal carcinoma screening pipeline in sub-Saharan Africa

Background Artificial intelligence for digital pathology may improve cancer detection and workflow efficiency in low- and middle-income countries, but most models are trained and validated in high-income settings, creating uncertainty about generalizability and real-world deployability under differing pre-analytic and infrastructure conditions. Objective To validate a locally optimized colorectal carcinoma screening pipeline, retrained on data from a local Kenyan cohort, and to assess its feasibility as a sensitivity-forward assistive pre-screening workflow. Methods Formalin-fixed, paraffin-embedded hematoxylin and eosin-stained slides from 136 biopsy-proven colonic adenocarcinoma cases and 20 normal controls from 2 Kenyan institutions were digitized at 40× using a Grundium Ocus scanner. Whole-slide images were partitioned at the case level. A feature-enrichment and candidate tile-selection step identified adenocarcinoma-rich regions. Candidate 512 × 512 RGB tiles were then adjudicated by a gastrointestinal pathologist into adenocarcinoma-containing and benign/non-neoplastic tile libraries. These expert-curated tiles were used to train and validate a supervised convolutional neural network classifier, which generated tile-level malignancy probabilities. Tile scores were aggregated into case-level predictions using Top-K pooling and positive-tile burden rules to support sensitivity-forward screening. Results Among 14,452 tiles from 70 quality-controlled cases and 19 controls, the model achieved strong tile-level discrimination (sensitivity 0.9548, specificity 0.9926, F1 score 0.9686, and AuROC curve 0.966); case-level Top-K aggregation achieved an AuROC of 1.0. Conclusion A locally optimized computational pathology screening pipeline, based on data from a local population, demonstrated robust colorectal adenocarcinoma detection in a Kenyan cohort. These findings suggest that existing machine-learning models can be adapted to local populations through retraining with locally derived data.