Improved Deep Learning for Cervical Cancer Screening
C. Sagar, Kavya Kavya, Anupama Bhan · 2023
Women all around the world are affected by cervical cancer, which is a serious public health issue. For cervical cancer to be successfully treated and managed, early identification is essential. Deep learning algorithms have considerable potential for enhancing cervical cancer detection precision. We compare several loss and activation functions for the classification of cervical cancer using the MobileNet architecture in this research. To train and test our algorithms, we used the Herlev dataset, which consists of 917 pictures from cervical biopsies. Six distinct loss functions were tested: dice loss, categorical Cross-entropy loss, Binary cross-entropy loss, Tversky loss, and Focal Tversky loss. In addition, we assessed the performance of the four activation functions sigmoid, exponential linear unit (ELU), and scaled exponential unit (SELU). Our results show that the combination of Focal Tversky loss loss function and SELU activation function achieved the highest classification accuracy of 98.72%. Our results show that the selection of loss and activation functions can significantly affect the MobileNet architecture's classification accuracy for cervical cancer.