- Abstract: Objectives: (i)Explore a new method for applying artificial intelligence to cervical cytology diagnosis. (ii)Realize an automatic detection system of positive cervical squamous epithelial cell. Methods: (i) The method can be divided into two phases: training and testing. (ii) For the training phases, first of all, we collect 500 cervical cell slides. Through scanning, annotating, and extracting patches, we get the training set, validation set and test set. Then, based on the Faster R-CNN (Faster Regions with CNN features) , we construct a neural network model for cell detection and classification. After continuous training, validation, analysis and tuning, we finally get a well trained neural network model. (iii) During the testing phase, we use the previously well trained neural model to predict the test set which consists of 100 whole slide images of cervical cytology. The model detects the five types of target cells at first, and then counts the number of cells in each category, finally, generates a diagnosis. Results: The positive precision rate on the validation set is 0.91. On the test set, for two-class problem,the accuracy is 0.78. For four-class problem, the accuracy is 0.70. Conclusion: Object detection technology has unique advantages in applying to cervical cytology. Through accurate detection and classification of various types of abnormal cells, as well as the statistics of each category, a comprehensive conclusion is made. This idea adopted in this study accords with doctors' traditional diagnosis process to a certain degree. Results show that the intelligent system realized with deep learning technology has the advantages of high speed, high consistency, and well diagnostic performance.
- Keywords: cervical cancer, cytology, object detection, Faster R-CNN
- Author Affiliation: Pathology division of Shenzhen Second People's Hospital，Semptian Co., Ltd. Machine Learning Lab