Combining machine learning and deep learning approaches to detect cervical cancer in cytology images

Abstract

Cervical cancer is the fourth most common cancer in women. When diagnosed early on, it is one of the most successfully treatable types of cancer. As such, screening tests are very effective as a prevention measure. These tests involve the analysis of microscopic fields of cytology samples which, when performed manually, is a very demanding task, requiring highly specialized laboratory technologists (cytotechs). Due to this, there has been a great interest in automating the overall screening process. Most of these computer-aided diagnosis systems subject the images from each sample to a set of steps, more notably focus and adequacy assessment, region of interest identification and respective classification. This work is focused on the last two stages, more specifically, the detection of abnormal regions and the classification of their abnormality level. The main approaches can be divided into two types: deep learning architectures and conventional machine learning models, both presenting their own set of advantages and disadvantages. This work explores the combination of both of these approaches in hybrid pipelines to minimize the problems of each one whilst taking advantage of the best they have to offer, ultimately contributing to a decision support system for cervical cancer diagnosis. More specifically, it is proposed a deep-learning approach for the detection of the regions of interest and respective bounding-box generation, followed by a simpler machine-learning model for their classification. Furthermore, a comparative analysis of different hybrid pipelines and algorithms will also be performed, aiming to support future research of similar solutions.