A fully-automated deep learning pipeline for cervical cancer classification

Abstract

Cervical cancer ranks the fourth most common cancer among females worldwide with roughly 528, 000 new cases yearly. Around 85% of the new cases occurred in less-developed countries. In these countries, the high fatality rate is mainly attributed to the lack of skilled medical staff and appropriate medical pre-screening procedures. Images capturing the cervical region, known as cervigrams, are the gold-standard for the basic evaluation of cervical cancer presence. Cervigrams have high inter-rater variability especially among less skilled medical specialists. In this paper, we develop a fully-automated pipeline for cervix detection and cervical cancer classification from cervigram images. The proposed pipeline consists of two pre-trained deep learning models for the automatic cervix detection and cervical tumor classification. The first model detects the cervix region 1000 times faster than state-of-the-art data-driven models while achieving a detection accuracy of 0.68 in terms of intersection of union (IoU) measure. Self-extracted features are used by the second model to classify the cervix tumors. These features are learned using two lightweight models based on convolutional neural networks (CNN). The proposed deep learning classifier outperforms existing models in terms of classification accuracy and speed. Our classifier is characterized by an area under the curve (AUC) score of 0.82 while classifying each cervix region 20 times faster. Finally, the pipeline accuracy, speed and lightweight architecture make it very appropriate for mobile phone deployment. Such deployment is expected to drastically enhance the early detection of cervical cancer in less-developed countries.