Transfer Learning for Classification of Uterine Cervix Images for Cervical Cancer Screening

摘要

Automated analysis of digital cervix images acquired during Visual Inspection with Acetic acid (VIA) is found to be of great help to aid the physicians to diagnose cervical cancer. Traditional classification methods require many features to distinguish between normal and abnormal cervix. Selection of distinct visual features which well represent the data and at the same time are capable of performing discriminative learning is complex. This problem can be overcome using deep learning approaches. Transfer learning is one of the deep learning approaches, which facilitates the use of a pre-trained network for a specific problem at hand. This paper presents a transfer learning using AlexNet, which is a pre-trained convolu-tional neural network, for classification of the cervix images into two classes namely negative and positive. This study used 2198 cervix images with 1090 belonging to negative class and 1108 to positive class. Our experiment using AlexNet for transfer learning achieved an accuracy of 0.934.