Medical Image Recognition Based On Regional Convolution Neural Network

Medical image can provide important clinical diagnosis information,and a lot of diagnosis depends on the recognition of medical image.In recent years,due to the continuous improvement of equipment performance,a large number of medical image data has been generated,which provides more abundant feature information.How to accurately and effectively identify useful information from a large number of medical images is of great significance for disease diagnosis.Regional convolution neural network(R-CNN)is widely used in medical image recognition.Therefore,this paper focuses on the medical image data of lung CT,and combines R-CNN to complete the recognition of pulmonary nodules.The purpose of this study is to assist doctors to detect and diagnose pulmonary nodules,reduce workload,reduce misdiagnosis rate and missed detection rate,and play an important role in the early diagnosis and treatment of lung cancer.The main contents of this paper are as follows:1.Preprocess the data set.In order to reduce the interference to the experiment,it is necessary to segment the lung parenchyma from the CT image.The threshold segmentation method is used to generate the binary image of CT image,and a series of operations such as corrosion,closure and normalization are carried out to extract the lung parenchyma,so as to make the data conform to the input standard of neural network.Chinese style.2.3D DPN-U Faster R-CNN model for candidate pulmonary nodules detection is proposed.The model is improved on the basis of 3D DPN-Unet as feature extraction network,while 3D DPN-Unet network is designed on the basis of Unet structure with 3D dual path network(DPN)as building block.Because pulmonary nodules belong to small target detection,the traditional fast r-cnn has poor detection effect on small-scale objects;the Unet backbone network captures multi-scale information through jumping connection,increases the sensitivity to nodule detection,and the 3D DPN network has compact structure,less parameters,and is easy to train,and can make full use of the 3D characteristics of CT data.Therefore,3D DPN and Unet are combined to design feature extraction network of Faster R-CNN detector.Experimental results show that the improved new detection network can enhance the accuracy of candidate nodule detection.Chinese style.3.3D Deep DPN-LGBM model is proposed for the classification of benign and malignant pulmonary nodules.The model is a combination of 3D Deep DPN network and light gradient boosting machine(light GBM)algorithm.For the classification of benign and malignant pulmonary nodules,we need to learn more abundant feature information of candidate nodules,which requires higher performance of classification model;while Deep DPN can extract more precise features,and light GBM has better classification performance under effective featuretraining.Therefore,3D deep DPN is designed to extract more advanced features by focusing only on the candidate pulmonary nodule area,and construct features with the introduced candidate nodule size and original pixel extra features.After providing complete and effective features,light GBM is trained to classify the benign and malignant candidate nodules.The experimental results show that the proposed classification framework can achieve more accurate classification of benign and malignant,and effectively reduce false-positive.