Diagnosing Cervical Cancer with Deep Learning

Abstract

Cervical cancer ranks as the fourth most prevalent cancer worldwide, affecting mostly the developing countries. However, early diagnosis can help facilitate the clinical management of the patient. The problem lies in the sparse presence of qualified and professional health cytotechnicians as compared to the number of people that need to be diagnosed. Computer-Aided Diagnostic system can be of a lot of help in making the diagnosis more accurate, reliable, faster and cheaper. Most of the existing algorithms require precise image segmentation to distinguish the cell. The traditional machine learning diagnostic system work similarly to the cytopathologists who rely on handcrafted morphological features such as nucleus area, nucleus-cytoplasm perimeter ratio, etc to determine the malignancy in a cell. However, our study uses Convolutional Neural Networks(CNN) which could potentially allow us to eliminate the computationally expensive tasks of segmentation and feature selection. Our results evidenced best accuracy scores of 96.25% for binary classification and 66.87% for seven class classification, which are comparable to the results achieved with established Machine Learning techniques. This study addresses the different aspects of training Deep networks on a publicly available cervical cancer database by Herlev Hospital. We also did a comparative investigation to establish the most suitable working hyperparameters, optimizers and classifiers for the dataset.