Stacking-based ensemble for cervical cancer detection

Cervical cancer remains a critical health challenge, and early detection is essential for effective treatment. This study introduces a robust approach for cervical cancer detection using an ensemble of pre-trained deep learning models DenseNet121, MobileNet, and Xception. Our model is fine-tuned with chosen layers trainable to enhance their performance in classifying cervical cancer images. The ensemble leverages stacking, at which the outputs of the base models are paired into meta-features, and a logistic regression model is used as a metaclassifier to make finalized predictions. Data pre-processing includes normalization augmentation to improve model generalization. A 5-fold cross-validation approach is employed to ensure comprehensive evaluation and robustness. Findings indicate that the stacking-based ensemble attains an overall accuracy of 0.9771, outperforming the individual models in classification accuracy and robustness. The average accuracies of an individual models are as follows: DenseNet121 (0.9491), MobileNet (0.9595), and Xception (0.9086). These findings indicate the effectiveness of integrating multiple deep learning models with stacking for medical image classification.