Thyroid Cytopathological Diagnosis Based On Weakly-supervised Multiple Instance Learning Using Deep Neural Networks

Currently,thyroid cancer has become one of the fastest growing cancers in the world.Early diagnosis of thyroid cancer can significantly reduce the mortality of patients.Pathological observation using thyroid puncture cytopathic WSI is the most effective way to conduct early diagnosis.However,in the thyroid puncture cytopathic WSI,the area of the effective diagnosis of follicular area is extremely small and scattered,and small lesions and early lesions are easily missed.With the rapid increase in the number of cases,pathologists are at increased risk of increased work pressure and misdiagnosis rate.Therefore,it is of great significance to study high-accuracy thyroid autocytopathic diagnosis techniques.According to the survey,the automatic thyroid cell pathology diagnosis at this stage is less,and requires a lot of accurate human marking,and the diagnostic accuracy is low.In this thesis,the actual clinical thyroid fine needle aspiration cytopathic WSI is used as experimental data.First,the key areas of the pathological image were located.The computer could not directly process the WSI and the blank background area in the WSI had a high proportion.The effective follicular cell area distribution was very small.In view of the above problems,this thesis proposes a standardized WSI processing flow,that is,noise reduction,Otsu threshold segmentation and cropping of WSI thumbnails to remove blank areas,and then fine segmentation of the follicular cell area of the original WSI residual image to filter out images of invalid diagnostic areas.After that,every WSI retains all small images containing follicular regions.At this time,the number is still large and the shape and size of the follicular regions are different.As follicular cells are overlapping and mixed,giving fine diagnosis of these images is very difficult.In order to solve this problem,a multi-instance diagnosis framework based on deep neural network is designed in this thesis.There is no need to mark the diagnosis results of each small picture.Each WSI is used to randomly extract several different images as a data packet and send it to the network.Based on the structure of VGG-Net,ResNet,and ResNeXt,the image feature extractor was proposed.Three different multi-instance pooling methods were introduced and designed for experiments,and the experimental results were verified.After that,in order to make full use of the effective feature information and improve the accuracy of diagnosis,the Attention mechanism for multiple example frameworks was introduced.Attention modules were added at the feature level and the example level,respectively.Subsequently,FPN-like structural modules were designed to improve feature extraction Performance.Through verification,you can find that these modules improve the accuracy of diagnosis and achieve the expected results.