HYBRID ENSEMBLE OF TRANSFER LEARNING AND CUSTOM CNN ARCHITECTURES FOR AUTOMATED CLASSIFICATION OF CERVICAL CANCER CELLS FROM PAP SMEAR IMAGES

Abstract

Background: The occurrence rate of cervical cancer has been increasing in recent years and the mortality rates because of it have increased rapidly and can be commonly seen in women across the world. This is greatly due to the fact that cervical cancer has no visible symptoms, or it is usually asymptomatic in its early stages. It is detected in the later stages when it has become malignant and is spreading and causing harm to other body systems. If it gets detected in early stages, treatment is possible and effective as compared to later stages. Screening for cervical cancer is usually done using conventional Pap Smear test in which cervical cells are examined for abnormalities using microscopic examinations. This process is cumbersome, time consuming and requires expert pathologists otherwise there will be a lot of variability and errors in the final results. To address these challenges, the present work aims to provide an automated and robust system using DTL and ensemble techniques for accurate classification of cervical cancer. Methodology: This study uses three pre-trained CNNs - ResNet152, EfficientNetV2 S, ConvNeXt-Base and a Custom CNN with residual blocks and CBAM attention mechanism to classify cervical cancer cells of SIPaKMeD dataset into 5-classes. A novel ensemble of these four base learners was proposed by averaging the predictions from each of the four models and then finally predicting one class from the dataset for a given input image. Result: For the four base learners the accuracy was 96.71%, 95.56%, 94.57% and 90.62% for ResNet152, EfficientNetV2-S, ConvNeXt-Base and Custom CNN respectively. The ensemble model in the current research gave a classification accuracy of 97.53% on the 5-class SIPaKMeD dataset, higher than those of individual base learners. Conclusion: The performance of proposed ensemble reflects its superiority over base learners and to some the previous studies as well. A method for classifying cervical cancer cells from Pap smear images is provided by the current research project. Deployment of such systems in real world clinical settings can help medical professionals for better treatment plans and also improve patient experience.