CERVIA: A Hybrid Deep Learning Framework With Explainable AI for Automated Cervical Cancer Classification From Pap Smear Images

Cervical cancer is a primary health concern worldwide. The World Health Organization (WHO) estimated that there were almost 350,000 deaths caused by the disease in 2020. Early detection by means of a Pap smear is vital. However, manual examination is prone to inter observer variability and is very time consuming. Deep learning techniques provide a potential solution. Hence, this research presents CERVIA, a hybrid deep learning network combining InceptionResNetV2 and MobileNetV3 architectures by feature fusion strategy to achieve higher discriminative power across NILM, LSIL, and HSIL categories. The problem of unbalanced classes is solved by Conditional Generative Adversarial Networks, so that the dataset becomes balanced for all the classes. Two explainable AI methods Local Interpretable Model agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) are used to give clear interpretability, with SHAP being more closely aligned with the diagnostic features. A clinical report generation component converts the predictions into the real world information. CERVIA has been able to get 97.1% accuracy, 96.7% precision, 97.1% recall, 96.9% F1 score, and 0.99 ROC AUC on the CRIC test set thus, performing far better than the most advanced methods. Independent validation with the Brown Multicellular ThinPrep Database was also able to confirm the resulting high accuracy of 96.8%, hence, generalization is very robust. The amalgamated framework sets up a comprehensive solution that a medical professional can employ in various clinics and hospitals, and other healthcare settings.