Submission ID 118251
| Issue/Objective | Cervical cancer remains a critical public health issue in Rwanda, where access to diagnostic tools and skilled healthcare providers is limited. This study explores the application of deep learning techniques for automated cervical cancer classification from histopathological images. |
|---|---|
| Methodology/Approach | Three deep learning models (ResNet50, EfficientNetB0, and DenseNet121) were used to classify 885 histopathological images from the University Teaching Hospital of Kigali (CHUK). Images underwent preprocessing, including resizing, RGB conversion, and augmentation to address class imbalance. Transfer learning with ImageNet pre-trained weights was applied, and models were evaluated using accuracy, sensitivity, specificity, and ROC-AUC metrics |
| Results | ResNet50 achieved the highest accuracy in binary classification (95%) with a ROC-AUC of 0.98, followed by EfficientNetB0 (91%, ROC-AUC: 0.92). DenseNet121 performed less effectively (74%, ROC-AUC: 0.85). Multi-class classification posed greater challenges, with ResNet50 attaining 65% accuracy, revealing limitations related to class imbalance and morphological similarities between cell types. EfficientNetB0 and DenseNet121 achieved 58% and 52% accuracy respectively in the multi-class task. |
| Discussion/Conclusion | ResNet50 shows strong potential for automated cervical cancer screening in resource-limited settings. While binary classification is reliable, multi-class classification needs improvement. Future efforts should focus on dataset expansion, optimized augmentation, and clinical validation to enhance model reliability and scalability |
| Presenters and affiliations | RUHUMURIZA Kaboyi Yves CENTRE FOR RESEARCH AND INSTITUTIONAL DEVELOPMENT (CRID) Limited UWAMALAYIKA Alice King Faisal Hospital Rwanda Foundation and Africa health sciences university uwase Nkusi Diana society for Family Health (SFH), Kigali, Rwanda |
