An Improved Deep Learning Based Cervical Cancer Detection Using a Median Filter Based Preprocessing

Abstract

Cervical cancer is one of the prevalent type of cancer among women although its treatment success is the highest when compared to other types of cancer once diagnosed. Automatic classification of cervical cancer is essential to accelerate the treatment process and increase the survival rate of the patients. Inadequate awareness, deficiency of medical opportunities, and expensive screening procedures increase the death rates. This common cancer is frequently screened by several imaging tests including Pap smear, cervicography and colposcopy. The decisions are made by the help of these tests, but structural complexities of cervical cells may complicate the decision. Recent developments in neural networks show remarkable achievements in disease diagnosis. Also, transfer learning draws the attention of most of the researchers because of its advantages. This paper presents a transfer learning based cervical cancer detection method for early diagnosis. Pap smear images were preprocessed using median filter before training the deep learning model in order to remove noise from the images for better classification. Cancerous and non-cancerous cervical cells are distinguished through pre-trained networks. Five popular pre-trained networks which are SqueezeNet, VGG-19, AlexNet, ResNet-50 and InceptionV3 have been utilized and compared for the problem. SqueezeNet achieved the best validation accuracy (96.90\%) when compared to other neural structures and this performance makes the proposed method the best among other unsupervised approaches in the literature for cervical cancer diagnosis. Additional experiments also proved the success of the proposed model for the classification of two similar classes, namely Parabasal and Metaplastic cells. The results demonstrate that the proposed approach can provide a confidential, cheap, and fast decision support system for cervical cancer diagnosis.

Topics & Concepts