Exploring Deep Convolution Neural Networks with Transfer Learning for Transformation Zone Type Prediction in Cervical Cancer

Cancer of the cervix is one of the most common types of women’s cancer. Cervical malignancy can be reasonably neutralized if distinguished in the initial stage. Classification of medical images is known to be a difficult problem for a number of reasons, but recent advancements in deep learning techniques have shown promise for such. In order to appropriately treat cervical cancer, making a right decision of a right type patient’s cervix type is necessary. The patient cervixes are classified as type 1, type 2, and type 3. We attempt to model a deep neural network by utilizing the fine-tuned transfer learning for classifying the cervical images. The model uses pre-computed activations of ResNet and Inceptionv3 which was pretrained on ImageNet Challenge. The dataset is taken from Kaggle challenge. It comprises 1481 training images, 512 test images, and 6734 additional images that we utilized for training model. Because of the less availability of the dataset, we apply image augmentation techniques. We experiment model with and without augmented data. The results show that model using the pretrained weights of ResNet performs better with top-down augmentations which results in 72% of accuracy on test set.