Automatic Cervical Cancer Screening Methods Based On Liquid-based Cytology Images

Cervical cancer screening based on liquid-based cytology im-ages is an important means of cervical cancer screening,which is of great significance for the prevention and control of cervical cancer.At present,cervical cancer screening in our country mainly relies on manual reading.Pathologists have a large workload,and the diagnosis results are limited by factors such as the experience and subjectivity of the readers,which can easily lead to misdiagnosis and missed diagnosis.With the rapid develop-ment of deep learning technology,computer-aided cervical cancer screen-ing has become possible,but the existing automatic cervical cancer screen-ing methods still need to be improved in detection accuracy and recognition accuracy.This paper proposes deep learning-based methods for automatic screening of cervical cytology smears from the cell level and smear level.In this paper,deep learning technology is used to automatically screen cervi-cal liquid-based cytology images from two aspects: abnormal cell detection and slide classification.The main work of this paper is as follows:Existing cervical cell detectors rely on single-region features to detect cells.This paper simulates the process of pathologists comprehensively comparing different cells to determine a single cell type during diagnosis,and proposes a cell detection model based on the relationship of region of interest(ROI).On the basis of Faster R-CNN,the model adds two modules based on the attention mechanism,the ROI-relationship attention module and the global ROI attention module.The two modules enrich Ro I features with intercellular region relation context and local cell region and global relation context for better detection results.The method in this paper is im-plemented based on the attention mechanism and requires only a few addi-tional parameters.The two modules are plug-and-play and can be deployed in any two-stage object detection method.On the cervical cell detection dataset collected in this paper,our method achieves an AP50 of 53.3% and a m AP of 30.4%,outperforming the baseline methods by 3.0% and 1.7%,respectively.The AP of our method exceeds all current cervical cell detec-tion methods and some classical general target detection methods.Cervical detectors can provide relatively accurate cell-level results,but when applied to slide classification,it will suffer from false positive results and lead to catastrophic errors in slide classification.Some existing slide classification methods set manually defined rules according to the results of cell detection or select some cell results for slide classification,which have poor classification effect,and have problems such as low utilization of whole slide images(WSI)image information and poor generalization.This paper proposes a Transformer-based slide classification method for the classification of cervical slide.The slide Transformer is carried out in two stages.The first stage divides the cervical WSI into local image blocks,and extracts the image block features through the local image block feature ex-traction network.In the second stage,all local image patch features in WSI are input into Transformer for slide classification.Our method makes full use of the information in WSI and can be trained and deployed end-to-end.On the cervical slice data set collected in this paper,our method achieves a sensitivity of 91.2%,a specificity of 93.8% and an accuracy of 93.5%.While ensuring high sensitivity,compared with the existing slice classifica-tion methods,the specificity and accuracy of our method have been greatly improved compared to the existing slide classification methods.