| Computer assisted cervical cell screening technology improves the screening efficiency of doctors and contributes to the popularization of cervical cancer screening.In recent years,with the development of deep learning,deep learning technology also began to be applied in the field of cervical cell screening,through training model to recommend suspicious cells,to assist doctors in screening.In deep learning,the training of cervical cell model is a process of learning doctors’ experience.Firstly,collect the label of cervical cancer cells from doctors,and train the model according to the label configuration data set.Then use the trained model to judge suspicious cervical cells and deliver the results to doctors for judgment and feedback.Finally,train the model again according to the feedback of doctors,and gradually improve the accuracy of the model.There are still some problems in the above process,which affect the fast iteration of the model: 1)there is a large amount of smear image data,and an image contains billions of pixels,so it is inefficient to use the model to infer smear image in a common way.2)The separation of each step leads to the difficulty of work handover.Especially between the model developer and the doctor,there is a lack of unified handover specification,which makes it very difficult to label and hand over the cervical lesion cells;3)The data management of each step is complex.For example,the collection of multi-source smear images and the management of multi category annotation in annotation lead to the slow iteration of the model.Aiming at the low efficiency of smear image model infering: 1)design read-write optimization algorithm to speed up the IO efficiency of smear image;2)The foreground segmentation algorithm is designed to remove the cell-free area of smear image and reduce the amount of calculation;3)A multithreading scheme is designed to allocate a reasonable number of threads for each process of model reasoning to maximize the utilization of computing resources;4)The structure of the model is optimized to speed up the reasoning speed of single image.Based on the above optimization algorithm,the inference performance of smear image by model is improved more than 5 times.This paper also studies the extension of smear image model reasoning,and designs a scheme to adapt to a variety of data formats of smear images and various types of models.Aiming at the separation of each step in the learning process and the problem of data management,an online learning system is designed.By designing the database of the system,the data management of each step of the learning process is completed;By combining with the database and designing the management interface for each step of the learning process,the online fast data handover of each step of the learning process is completed.In the system,online statistical annotation can be carried out to complete the configuration of data sets;online model training;online infer cervical smear image by model;online annotation of the diseased cells recommended by the model.On the basis of method research,the online learning system of cervical lesion cells was completed.The system iterates the model of multi-source cervical smear images,and improves the accuracy of the model from 0.86 to 0.97,which proves the effectiveness of the system and provides a powerful tool for model developers and annotation doctors. |
PDF Full Text Request