Abnormal Region Detection Of Cervical Cytology Images Based On Convolutional Neural Network And Transfer Learning

Cervical cancer is the most common malignant tumor in gynecology,and the only cancer with definite etiology and early prevention and treatment.Due to the large population,automated slide reading has become a necessary condition to meet the demands of the regular cervical cancer screening for women in China.It is urgent to develop a computer-aided slide reading system and one of the most critical steps is to locate abnormal regions from the massive cells of cervical cytology images quickly and accurately.This paper proposes an abnormal region detection algorithm for cervical cytology images based on convolutional neural network and transfer learning.Firstly,a framework for detecting abnormal regions of cervical cytology images based on convolutional neural network is designed.The framework is divided into three parts: feature extraction subnetwork,region coarse detection subnetwork and region refined detection subnetwork.The feature extraction subnetwork is mainly used to extract the image features.According to the characteristics of small target area and scale sensitivity of the cervical cell images,a feature extraction network capable of accurately extracting high-level semantic information and low-level detailed information of cervical cell images is designed and implemented.The region coarse detection subnetwork is used to generate as many target candidate regions as possible from the original image.The region refined detection subnetwork is constructed by a full convolutional network,which can carry out accurate reclassification and position regression of candidate regions.A scale-sensitive region of interest pooling layer is proposed in the region fine detection network for the scale-sensitive characteristic of cervical abnormal regions.In addition,an improved evaluation scheme is designed to emphasize accurate positioning in the detection of abnormal regions of cervical cytology images.In order to solve the problem of difference in cervical cytology image data between different batches,this paper also proposes a cervical abnormal region detection algorithm based on transfer learning.A multi-level transfer learning method is used to achieve domain adaptation,including three different levels: image level,region level,and region of interest level.At the same time,the attention mechanism is added in the region level domain adaptation module to achieve differential domain adaptation according to the degree of transfer of different regions of the image.This paper designs and completes the detection of abnormal region of cervical cell images based on object detection and transfer learning respectively.In the experimental part of the convolutional neural network,comparative experiments of different feature extraction subnetworks and the experiments to verify different thresholds and effects of data enhancement are designed.In the experimental part of the transfer learning module,the experimental results comparison experiment and the ablation experiment for the multi-level domain adaptation module are designed and implemented.The experimental results show that the proposed algorithm can meet the needs of clinical cervical cytology examination,and the performance is improved compared with other algorithms.