Research On The Auxiliary Diagnosis Of Pleomorphic Adenoma Based On Deep Learning

With the continuous development of computer software and hardware,computerrelated methods are constantly applied in various fields and people’s life scenes,among which intelligent medicine is one of the most important application fields.In view of the imbalance between medical supply and demand,the lack of professional imaging doctors and the uneven professional abilities of doctors,intelligent medicine provides an effective solution.At present,computer aided diagnosis technology has been developed and applied in lung,breast,colorectal and prostate imaging.Pleomorphic adenoma is one of the most common types of epithelial tumors of salivary glands with high malignancy.According to statistics,its incidence is increasing year by year.Although pleomorphic adenoma grows slowly,it has a high probability of recurrence after surgery and multiple recurrence may further lead to malignant changes,reduce people’s quality of life and endanger life and health.The diagnosis can be obtained through a variety of examination methods.The doctor can make a rough diagnosis based on the medical history and the patient’s clinical symptoms.However,the final diagnosis must rely on the pathological diagnosis.At present,the pathological diagnosis of pleomorphic adenoma is completely dependent on artificial visual discrimination and observation,and its pathological images are characterized by tissue pleomorphism.The diagnosis requires doctors to have sufficient professional knowledge,spend a lot of time on observation and processing,and may cause visual fatigue,leading to misdiagnosis and missed diagnosis and delay of the patient’s condition.In order to make up for the above problems in this field,this thesis uses computer technology to provide certain auxiliary work for diagnosis or auxiliary information for doctors’ diagnosis,and proposes an auxiliary diagnosis algorithm for pleomorphic adenoma based on deep learning related methods.It realizes qualitative diagnosis based on pathological images through a two-stage model to achieve auxiliary purpose.In this thesis,by learning and referring to Dense Net model and decision tree model,the first stage model is constructed to realize the imagebased instance segmentation Den SE2 Net.In the second stage,the probability-based diagnostic decision model is constructed and applied to the auxiliary diagnosis of pleomorphic adenoma.The main work is as follows:(1)The pathological section data set of pleomorphic adenoma was constructed.The original data were collected from pathological sections of pleomorphic adenoma patients in Stomatological Hospital of Jilin University,200 times larger under the lens.After learning relevant pathological knowledge,images were preprocessed and converted into various tissue sections.The four types of key tissues were preprocessed,and the data set suitable for the model in this thesis was screened and constructed,with a total of 48,230 pieces of small section data.(2)A two-stage diagnostic model for pleomorphic adenoma was proposed.In the first stage,Den SE2 Net model is proposed to realize image-based instance segmentation.Based on dense linking,the model made full use of the low-level features of tissue cells in pathological images,learned the Bottle2 Neck structure of Res2 Net to improve the backbone network of the model,and then improved the sensitivity of the model to multi-scale structures.The fusion of SE Net structure enabled the model to adjust the weight of channel features adaptively.Multiple experiments were designed to verify the proposed model.Compared with the current advanced network model,the accuracy of the proposed model reached 0.977,which fully verified its effectiveness and advancement.(3)The second stage model is trained to realize probability-based diagnostic decisions.Based on the organization category and its probability score obtained from the first-stage model,the CART decision tree model is selected for the probabilitybased diagnostic decision in the second stage,and the inference accuracy of 1.0 is obtained through experimental verification.In addition,manual management branches are provided and information is provided for cases that may be in doubt.(4)Verify the universality of the model in this thesis.The performance of the proposed model Den SE2 Net was verified by the open source dataset subtype classification task.The experimental results showed that the classification accuracy reached 0.992,indicating that the proposed model has good applicability to the classification task of medical images.