Multi-branch deep learning architecture for improved colposcopy image classification

Abstract

Cervical cancer is still the major cause of mortality among women. However, substantial variability in lesion appearance, illumination, and tissue characteristics poses significant challenges to accurate and reliable classification. In this study, we propose a novel multi-branch deep learning framework that integrates complementary feature representations, attention mechanisms, and ensemble learning to enhance diagnostic performance. EfficientNetB0 and MobileNetV2 are employed as parallel feature extractors to capture rich and diverse visual features from colposcopy images, which are subsequently fused and classified using a soft-voting ensemble of Logistic Regression, XGBoost, and CatBoost classifiers. Multitask learning is incorporated to address lesion-specific classification objectives, while hyperparameter optimization and k-fold cross-validation are applied to ensure robustness and generalization. Experimental results demonstrate that the proposed framework achieves a validation accuracy of 99.85% under five-fold cross-validation, outperforming individual model configurations and exhibiting strong generalization capability. These findings suggest that advanced hybrid deep learning architectures can provide a reliable and scalable approach for colposcopy-based cervical cancer detection, with potential implications for supporting clinical decision-making in future applications.