Abnormal Region Detection Of Cervical Cytology Images Based On Object Detection And Adversarial Transfer Learning

Cervical cancer is one of the most common and increasing malignant tumors in the diagnosis of female diseases.It is of great significance to diagnose and treat cervical cancer early.Cytological examination is the most commonly used technique for early diagnosis of cervical cancer.At present,cytological examination is completed by medical or laboratory professionals through microscope observation in China,but manual examination of slides may be unstable.Using computer vision technology to assist examination can greatly improve efficiency and accuracy.A key step is to quickly locate abnormal areas from a large number of cervical cytological images.This thesis proposes the algorithm for locating abnormal regions in cervical cell images based on object detection and adversarial transfer learning.Firstly,a single-stage,anchor-free framework for abnormal region localization of cervical cytology images based on feature fusion and cascaded attention was proposed.Firstly,the residual feature extraction network is used to extract image backbone features.Secondly,a pyramid feature network based on channel attention is designed,and channel attention without channel information attenuation is introduced to adaptively adjust the channel feature response.In order to improve the detection performance of small targets in cervical cytology,the upper-level features were fused to lowerlevel features to obtain lower-level features with stronger expression ability,then the features with stronger expression ability are obtained.Finally,a shared detection head based on cascaded channel attention and spatial attention is designed to enhance the weight of feature points and channel.In addition,an evaluation branch of the quality of the bounding box is designed to quantitatively analyze the quality of the prediction boxes and reduce the number of negative sample boxes so as to train the bounding box classifier with better performance.To solve the image distribution differences caused by the collection process of different batches of datasets,this thesis also proposes a method for locating abnormal regions in cervical cytological images based on adversarial transfer learning.Adopts adversarial transfer learning method,will be based on the object detection of abnormal cervical cytological image location framework of residual feature extraction and feature fusion network joint generator,share the detection head as a detector,design a multi-level network domain classification,characteristics of generator extracted a multi-level domain classification,to study domain invariant features for detecting,Spectral normalization method is added to the convolutional layer of each domain classification network to solve the unstable factors in the process of confrontation training.In this thesis,we designed and completed the localization experiment of abnormal region of cervical cell image based on object detection and the localization experiment of abnormal region of cervical cell image based on adversarial transfer learning.Firstly,the dataset is augmented in various ways to avoid over-fitting.Experimental results show that the AP of the proposed object detection method is better than that of the mainstream object detection method on the test set.In addition,the ablation experiment proves the effectiveness of the proposed module.In addition,the adversarial transfer learning,without transfer learning and mainstream adversarial transfer learning method are compared experimentally to prove that based on the good performance of adversarial transfer learning method,ablation experiments are carried out to prove the rationality of the multi-level domain classification structure designed in this paper.