An Evaluative Investigation of Deep Learning Models by Utilizing Transfer Learning and Fine-Tuning for Cervical Cancer Screening of Whole Slide Pap-Smear Images

Deep learning (DL) is a prominent tool utilized today in many applications across many industries, including the healthcare realm. DL methods can manage several problems that traditional artificial intelligence (AI) methods find challenging. In this paper, we analyzed the performance of nine prevalent DL models i.e. VGG-16, DenseNet-121, ResNet50, VGG-19, DenseNet-169, Xception, EfficientNetB0, InceptionV3, and ResNet-152 pre-trained on ImageNet dataset for cervical cancer screening. These previously trained models are fine-tuned by utilizing transfer learning (TL) for 5-class and 2-class classification of whole slide pap-smear images (WSI). Two-step data augmentation is being used for the preprocessing of data to enhance the efficacy and robustness of classifiers by increasing the amount of the training data and reducing overfitting. Among the aforementioned DL methods, VGG-16 performs best among all with an accuracy of 94.89% for 5-class and 97.16% for 2-class classification.