Atypical Squamous Cells Detection Based On Deep Learning

Cervical cancer is the most common gynecological malignant tumor in the world,with a very high mortality rate.Thinprep Cytologic Test(TCT)is the basic method for cervical cancer screening.Pathologists observe squamous epithelial cells shed from the cervix under a microscope to see if there are abnormal squamous epithelial cells for diagnosis.The detection rate of cervical cancer by TCT is 100%,and some precancerous lesions and microbial infections can also be found.At present,there are only about10,000 pathologists in China,and the training cycle is long,with a huge demand gap.The work of this thesis mainly includes the following three aspects:1.Design and realize the algorithm with "re-blurring" approach for digital pathological image clarity evaluation.It solves the problem that the algorithm based on traditional methods in digital image processing(image variance,Laplace gradient,image energy function,Vollath,etc.)are not able to effectively distinguish between a clear image with less cells and a blurred image with more cells.Based on the “re-blurring”algorithm,the detection accuracy index AUC(Area Under Curve)with 2000 digital pathological images can reach 98.78%,12.593% higher than that of image variance method,it can automatically pick out the blurred images in the model training and inference process.2.Training positive cells detection model through pathologist-labeled digital pathological images.Based on the improved model of Faster R-CNN network structure,Deformable Convoluational Network(DCN)and Feature Pyramid Network(FPN)are introduced.The experimental results show that the three models(including the model with FPN,the model with DCN,and the model with both FPN and DCN)all can converge quickly.3.Model evaluation and clinical data verification.The model with both FPN and DCN is tested and evaluated thoroughly.The result of m AP(mean Average Precision)on clinical data is 0.29(the m AP used in actual clinical practice is 0.32).Comparing with other models,it is more effective and more accurate,has basically satisfied the needs of assisting pathologist in diagnosis,and proved that FPN and DCN have played critical role in small object detection and extraction of irregular object feature.