Automated Pap Smear Cervical Cancer Detection Based on Multiscale Convolutional Neural Network

Abstract

Cervical cancer is one deadly common gynecological malignant tumor, early accurate screening detection can save lives. However, the large-scale manual screening is limited by the few experienced cytologists, which is difficult to change under the existing conditions. In this paper, we propose an automatic detection framework based on deep learning to improve the detection accuracy and efficiency of cervical cancer cells. Different from the traditional convolutional neural network with a top-down structure, we design an isomorphic dual-branch residual-like structure as a multi-scale network. The multi-scale network can increase the range of the receptive field with each network layer at a fine-grained level to accurately learn the representation information of pathological images. In addition, we also perform ablation studies on different widths, residual block types, and patch extraction methods in the backbone network. The experimental results show that our framework has achieved an AP of 89.7% in cervical cancer cell detection, which is superior to other existing detection frameworks.