Classification And Segmentation Of Pathological Images Of Oral Squamous Cell Carcinoma Based On Convolutional Neural Network

Oral cancer is the general name of malignant tumors which occurring in the human oral cavity,including the gums,tongue,oropharynx,and alveolar mucosa and so on.Among all the cases of oral cancer,a large proportion are squamous cell carcinomas,and biopsy is an important way to diagnose oral cancer.Currently,the clinical diagnosis of oral squamous cell carcinoma is mainly made by observing the digital pathological images of oral cavity.The study on the automatic diagnosis of oral squamous cell carcinoma is conducive to the development of the computer-aided diagnosis system for oral cavity,and can reduce the working intensity of doctors,which has important clinical application value.As the increasing computing power of hardware computing devices like GPU,the convolution neural network based on deep learning has been widely concerned because of it’s convenience of automaic feature extraction and high recognition accuracy.In the actual clinical diagnosis,the qualitative diagnosis of oral squamous cell carcinoma is usually made based on the image characteristics of pathological sections of oral cavity first,and then according to the specific situation,the cancerous area in the pathological image which is confirmed oral squamous cell carcinoma should be located.Based on the requirements of actual clinical diagnosis,this thesis designed a classification diagnosis method of oral pathological section images based on convolutional neural network,and compared the output effects of relevant image semantic segmentation algorithm on the images of oral squamous cell carcinoma pathological section.Experimentally verified,the accurate prediction results of oral pathological section images and the segmentation and localization results of cancerous regions in the confirmed positive images were obtained.The main contents of this thesis are as follows:(1)Desensitized images of pathological sections of the oral cavity were collected which have been classified by senior pathologists.And the cancerous area of the images which has been classified as cancerous has been marked.After image preprocessing,the image data set of oral cytopathic sections was established.In the end,2850 normal cell plots and 2900 squamous cell carcinoma plots which containing cancerous areas were taken part in the image classification and diagnosis network training,and 16500 pathological plots of oral squamous cell carcinoma with gold standard were participated in the image segmentation network training.(2)The deep convolutional neural network based on the graph block was trained,with a small amount of data,the classification and diagnosis of normal or cancerous oral section images were completed,with a classification accuracy of 98.46%.(3)In this thesis,Res Net and Dense Net are used as encoders in the encod-decode structure respectively to verify the effectiveness of dense connection.U-net and UNet++structures were used respectively to train the segmentation task of digital pathological section images of oral squamous cell carcinoma,and the output results were analyzed and compared.The experiment showed that the UNet++ structure using Dense Net as the encoder obtained the best results,which was close to the gold standard,and could basically meet the needs of clinical cancer region positioning diagnosis.