| Pleomorphic adenoma is a highly malignant tumor,which often occurs in the mouth,jaw and face.Pleomorphic adenoma,also known as mixed tumor,is a parotid gland tumor divided into glandular duct,mucus tissue,cartilage tissue and myoepithelial cell tissue.Because pleomorphic adenoma is generally asymptomatic,painless and relatively slow growth,most of the masses are found by casual physical examination,and the course of disease is as long as more than ten years,which seriously affects human life and health,so the treatment of pleomorphic adenoma has become a problem of concern.At this stage,the pathological section images of pleomorphic adenoma are obtained through electronic computed tomography（CT）,ultrasonic diagnosis（Bultrasound）and magnetic resonance imaging（MRI）.Doctors can make a general diagnosis according to the medical history and clinical manifestations.However,different types of pleomorphic adenomas are not easy to distinguish.Strict professional knowledge is required in the diagnosis process,which takes a lot of time.Moreover,professional doctors engaged in the same work for a long time will have visual fatigue in diagnosis,which will lead to wrong or missing diagnosis of masses.Therefore,this paper proposes a detection and classification algorithm of pleomorphic adenoma based on convolution neural network to solve the problems of missed diagnosis,misdiagnosis and time-consuming in the process of manual diagnosis,so as to achieve the purpose of auxiliary diagnosis.By learning from the Mask RCNN model,this paper proposes an improved Mask RCNN model for the detection and classification of pleomorphic adenomas.The main work is as follows:（1）The data set of pleomorphic adenoma was established.The pathological images of pleomorphic adenoma from patients in stomatological hospital of Jilin University were collected and magnified 200 times by microscope.A total of 2552 images were obtained.By learning relevant pathological knowledge,according to the characteristics of different pleomorphic adenomas,the images are labeled and preprocessed,and different tumor cells are classified,so as to construct the data set required by this algorithm.（2）To determine the detection scheme of pleomorphic adenoma.This paper analyzes the classification reasoning and prediction of pleomorphic adenoma,focuses on the advantages and disadvantages of various existing target detection algorithms and convolutional neural network,and determines the improved pleomorphic adenoma detection scheme based on Mask RCNN.（3）To design a method to detect the characteristics of pleomorphic adenoma.The backbone network and feature fusion are improved,Res2 Net is used as the backbone network for feature extraction,and FPN is changed to FPG,so that complex hierarchical features can be learned across layers.（4）Design the experiment and analyze the results.A total of 10208 labeled images and 139 unlabeled images were obtained through image preprocessing,of which 75%of the labeled images were used as the training set,25% as the verification set,and the unlabeled images were used as the test set.At the same time,typical target detection algorithms are used for comparative experiments,and ACC and m AP are used as evaluation indicators to pay more attention to ACC.The experimental results show that the accuracy of this method is 95.45% and the m AP is 11.9%.The detection effect of accuracy is better than other target detection algorithms.（5）Verify the general applicability of this algorithm.Experiments were performed using the breast cancer dataset.The dataset was selected as an open Breakhis,marking17 images of ductal carcinoma,51 lobular cancer images,and 28 papillary carcinoma images as data sets.The trained neural network model based on improved Mask RCNN polymorphic adenoma was used as a pre training neural network model for breast cancer detection.The accuracy of detection and classification has reached 97.22%,so this algorithm is adaptive to the detection and classification of medical images. |
PDF Full Text Request