Research On Cervical Cancer Detection Algorithm Based On Deep Learning

Cervical cancer is one of the common gynecological malignant tumors.Regular screening and timely treatment can effectively reduce the mortality of women suffering from cervical cancer.Cervical cytology screening method requires pathologists to screen out abnormal cells from hundreds of thousands of cells on cervical exfoliated cell smear and make a diagnosis of the degree of cervical lesions.With the increase of reading amount,manual reading method is easy to be misdiagnosed because of doctor’s visual fatigue.Automatic detection of cervical cancer cells can assist doctors in diagnosis and improve the efficiency and accuracy of screening.For the existing cervical cancer cell detection algorithms,further research needs to be carried out on how to mine difficult negative samples in negative images and how to sample more representative candidate frame samples,besides the real-time detection on whole slide image(WSI)is insufficient.In this thesis,the algorithm of cervical cancer cell detection based on deep learning is studied.The algorithm is optimized in difficult negative sample sampling and region proposal mining,and a lightweight cervical cancer cell detection model is designed to solve the problem of insufficient real-time performance of the algorithm.The main research contents and achievements are as follows.The original data set of cervical cancer is analyzed and processed.For the problem that there exists a large number of negative WSIs without labels,a region of interst(Ro I)sampling algorithm is proposed to generate negative Ro I from cervical cell negative WSIs cancer with non-smear areas,which improves the efficiency of negative Ro I sampling.To solve the problem that GPU can’t accommodate Ultra-high resolution input image,the sliding window clipping method is used to crop Ro I and generate a patch dataset,which lays a research foundation for subsequent experiments and algorithm.A cervical cancer cell detection algorithm based on mining difficult negative samples is proposed.For the problem that the model will produce a large number of false positive false detections in the negative Ro I and the negative area of the positive Ro I,a virtual class generation algorithm is proposed to mine the difficult negative samples of the negative patch,which significantly reduces the false detections of the model in the negative Ro I and the negative area of the positive Ro I.The accuracy of Retina Net,Faster R-CNN and Cascade R-CNN is improved by 5.61%,5.85% and 7.54% respectively after using the algorithm of mining difficult negative samples.A cervical cancer cell detection algorithm based on hybrid sampling is proposed.For the problem that there exists a large number of indistinguishable background samples and noises such as uneven cell staining,broken cells,and other impurities in cervical cancer cell images,a hybrid sampling algorithm is proposed to sample region proposals,in this way a more representative sample distribution of region proposals is obtained.For the problem that the cell clumps and the cells in the cervical cancer cell image are detected at the same time,a non maximum suppression algorithm based on Intersection over Other(Io O)is proposed,which obviously reduces the cells falsely detected in the cell clump.In the cervical cancer dataset,the precision of Faster R-CNN and Cascade R-CNN is improved by1.92% and 1.89% respectively,and the precision of Retina Net is improved by 0.47% by the non maximum suppression algorithm based on Io O.An efficient cervical cancer cell detection network is designed.For the problem of modular redundancy in cervical cancer cell detection algorithm based on deep learning,a model compression algorithm is adopted to prune the feature extraction network of Retina Net,and a channel adjustment strategy is proposed to reduce the computation and parameters of the feature pyramid and the detection head network,in order to design an efficient network.Compared with Retina Net,the precision of the designed algorithm is improved by 1.74%,and the speed reaches 33.7FPS.