Abstract / Summary
Abstract Background and Objective Cervical cancer remains a major cause of morbidity and mortality in Sub-Saharan Africa (SSA), where access to pathology expertise, screening, tissue diagnosis and digital pathology infrastructure is uneven. Histopathological examination remains central to confirming cervical neoplasia, grading lesions and guiding management. Artificial intelligence (AI) offers opportunities to support pathologists by automating image segmentation and classification, but evidence on clinical generalisability and deployment in SSA is fragmented. This systematic review synthesizes AI techniques for cervical histopathology segmentation and classification and examines barriers to translation into routine practice in SSA. Methods A structured systematic evidence synthesis was prepared and reported with reference to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020). Contemporary evidence was identified from PubMed/MEDLINE, Scopus, Web of Science, IEEE Xplore and Google Scholar searches and from reference chaining of relevant systematic reviews. Eligible evidence addressed cervical histopathology or closely related cervical image-analysis tasks involving segmentation, feature learning, classification, whole-slide analysis, clinical validation or deployment. The synthesis focused on 49 substantive publications spanning classical image processing, machine learning, CNNs, transfer learning, U-Net-family and other segmentation methods, attention/transformer models, whole-slide analysis, African studies and digital-pathology implementation. Results The evidence shows a transition from thresholding, clustering, watershed, active contours and handcrafted morphological/texture features toward CNN-based segmentation and classification, transfer learning, multiple-instance learning, attention models and transformer/foundation-model approaches. Early studies were frequently based on small, pre-segmented public datasets, whereas newer studies increasingly use histology images and whole-slide images. African evidence remains limited but is expanding. An Ethiopian study used 915 histopathology images and EfficientNetB0, reporting 94.5% test accuracy for cervical cancer classification. A 2026 Rwandan study used 885 patient-level H&E cervical biopsy images and reported EfficientNetB0 accuracy of 0.99 with ROC-AUC 0.99 on a held-out test set. A Ugandan study used 5,966 cervical histopathology images acquired with smartphone-assisted microscopy and DeepLabV3 + with a ResNet34 encoder for 21-feature segmentation, reporting validation mIoU of approximately 75.8% and Dice of approximately 93.1%. Conclusion AI has substantial potential to augment cervical histopathology in SSA, but reported algorithmic performance should not be interpreted as evidence of clinical readiness. The dominant evidence gaps are external validation, prospective evaluation, multi-institutional datasets, standardised staining and imaging, transparent reporting, calibration, explainability, workflow integration and implementation economics. Future research should prioritize locally representative datasets, lightweight and edge-capable models, patient-level validation, human-in-the-loop decision support and implementation studies conducted within African pathology workflows.