Early Detection and Classification of Cervical Cancer Cells in Women: A Comparative Study of Transfer Learning Models

Cervical cancer remains a significant health challenge globally, especially in developing countries where it accounts for about 15% of all female cancers. Despite the availability of various diagnostic methods, the Pap-smear test is the most widely adopted for early detection. However, the manual analysis of Pap-smear images is prone to errors due to the high similarity between cell categories, leading to misdiagnosis and delayed treatment. This study aims to enhance cervical cancer detection using advanced deep learning techniques, specifically transfer learning models. We conduct a comprehensive analysis of four pre-trained CNN architectures-DenseNet201, Xception, InceptionResNet-V2, and MobileNet-V2-evaluating their performance on a publicly available dataset SIPaKMeD, which includes 4,049 Pap-smear images categorized into five classes. The models were assessed based on accuracy, precision, recall, and F1-score. Our results indicate that InceptionResNet-V2 outperforms the others with the highest accuracy, while MobileNet-V2 and Xception offer a good balance between performance and computational efficiency. The study demonstrates that transfer learning and fine-tuning significantly improve the classification of cervical cancer cells, even with limited labeled data. This signifies the great potential of transfer learning methods in strengthening early detection strategies for cervical cancer in women, which may revolutionize medical imaging and diagnostics. These promising findings provide valuable insights into the use of advanced deep learning techniques in cancer research and pave the way for future exploration and integration of such methods into clinical practice, thereby improving efforts to overcome cervical cancer and enhance healthcare outcomes globally.