Cervical Cancer Detection: A Comprehensive Evaluation of CNN Models, Vision Transformer Approaches, and Fusion Strategies

Abstract

Cervical cancer, a malignant tumor arising from the cervix, poses a significant health risk to women worldwide. Early detection plays a pivotal role in improving patient outcomes by enabling timely intervention and effective treatment. This paper explores three distinct strategies for enhancing cervical cancer detection. Firstly, four convolutional neural network (CNN) models - EfficientNetB, DenseNet, Xception, and ResNet50 - are evaluated across performance metrics including accuracy, recall, F1-score, and precision. Results indicate accuracy values ranging from 0.65 to 0.78, reflecting the challenges posed by low-resolution imagery. Secondly, the ViT-Cerv methodology is introduced, leveraging ViT architecture, trained from scratch on the SkipMed dataset. The ViT-Cerv model achieves an ACC of 0.94, demonstrating promise in cervical cancer classification despite the dataset's low resolution. Finally, the HViT-Cerv hybrid model, synergizing the strengths of both CNN models and the ViT-Cerv architecture through fusion, demonstrates enhanced accuracy ranging from 0.97 to 0.99. This underscores the efficacy of the fusion approach in achieving superior classification capabilities, showcasing a notable advancement in accuracy over individual models. A comprehensive comparison with state-of-the-art models highlights the significance of the proposed approaches in cervical cancer detection, particularly in overcoming challenges posed by low-resolution datasets.