부산대학교 도서관

Cervical cancer occurs mainly from a deterioration of pre-cancerous lesions and represents a global threat among females, ranking fourth for both incidence and mortality among cancers. Prevention is able to significantly diminish this concerning issue. In that sense, a valuable support for physicians could be a tool that is able to automatically classify images from colposcopy. In this paper, two methods based on the use of deep neural networks have been implemented and tested for the automatic classification of colposcopy images. One of the methods consisted in constructing a Convolutional Neural Network, specific for image classification, which obtained a training accuracy of 41% and a validation of 31%. Another method was implemented to improve the previous results, a Residual Neural Network composed of 50 hidden layers. This second implementation achieved a training accuracy of 93% and a validation of 60%, with validation costs similar to the first method. These preliminary results suggest that deep learning could be applied successfully to the automated classification of colposcopy images.