A Robust and Explainable Deep Learning Method for Cervical Cancer Screening

The cervix forms the lower part of the uterus and marks the passage between the uterine body and the vagina. The cervix is crossed by a fusiform canal which is called the cervical canal. Cervical cancer is a disease caused by the uncontrolled multiplication of some cells located at the level of the squamous-columnar junction which transform into malignant cells. It represents the leading cause of death from gynecological cancer in the world and almost half of the cases are recorded among women aged between 35 and 55 years. Early diagnosis is important to treat the disease in its early stage when the chances of recovery are highest. In this paper, we propose a method aimed to detect the presence of cancerous cells in cervix histological images for screening purposes. The distinct features of the proposed method rely on the explainability, thanks to the adoption of the activation map to provide a localization behind a certain prediction, and in the robustness, of using different activation maps to visually confirm the location of cancer cells. In this way the pathologist and the medical doctor can trust the prediction, considering that the localization of the cancerous cells is confirmed by two different activation map algorithms. Five different deep learning models are evaluated (i.e., Inception, ResNet20, DenseNet, MobileNet, and a Convolutional Neural Network developed by authors), obtaining an accuracy ranging from 0.78 to 0.83. We discuss the model effectiveness for the screening of cervical cancer cells not only based on quantitative results (i.e. how many pathological images they can correctly classify) but also based on qualitative results, i.e. by considering the quality of explainability and on the robustness of predictions.