Classification of Cervical Cancer from Pap Smear Images Using Deep Learning: A Comparison of Transfer Learning Models

Abstract

Abstract Purpose Globally, cervical cancer is the most diagnosed type of cancer, and it is particularly prevalent among women in developing countries. The Pap smear is an essential diagnostic tool for detecting abnormal cells associated with cervical cancer. It is possible to significantly reduce cervical cancer deaths if detected and treated early. The manual screening process, however, results in a high percentage of false positives because of human error, which results in unnecessary treatment and anxiety for the patient. Therefore, it is imperative to develop a screening method that is more accurate and efficient to reduce false positives. To overcome this problem, automated screening methods have been proposed, such as computer-aided diagnosis (CAD), which can provide a more accurate and efficient diagnosis.Design/methodology/approach In this regard, this paper uses Deep Transfer Learning (DTL) models to classify single-cell pap smear images. Several pre-trained DTL models have been evaluated, including VGG16, VGG19, ResNet50, ResNet50V2, ResNet101, ResNet101V2, ResNet152, ResNet152V2, DenseNet121, DenseNet169, DenseNet201, MobileNet, XceptionNet, InceptionV3, and InceptionResNetV2, to determine which model is the most effective for detecting cervical cancer.Findings Based on a comparison of the results, ResNet101 and ResNet50 performed best in terms of accuracy and precision. In terms of classification accuracy, ResNet101 was the most reliable model with a score of 95.56 percent, whereas ResNet50 was the second most accurate model with a score of 91.19%. Our findings indicate that DTL models are suitable for automating cervical cancer screening, providing more accurate and efficient results than manual screening.Practical implications These models provide cytologists with valuable insights into cervix abnormalities and a reliable and efficient method for analysing and interpreting pap smear images.Research implications Due to the advancement of deep transfer learning, it has become possible to accurately classify single-cell pap smear images, which is crucial for detecting cervical cancer. Furthermore, the novice researcher can consult the reference paper to determine which transfer learning model is most suitable for their analysis of the Herlev dataset.Originality/value The proposed model using ResNet101 maximizes classification accuracy when compared to VGG16, VGG19, ResNet50, ResNet50V2, ResNet101V2, ResNet152, ResNet152V2, MobileNet, XceptionNet, InceptionV3, and InceptionResNetV2. To validate the results, confusion matrices are computed for each model. An original contribution of the paper is to present 16 deep transfer learning models for the classification of cervical cancers based on the Herlev dataset.