Research On Medical Image Segmentation And Recognition Algorithm Based On Deep Learning

With the development and advancement of medical imaging technology,doctors have started to make extensive use of medical imaging data to achieve segmentation,extraction,3D reconstruction and classification identification of human tissues or lesions.Such technology has gradually evolved into a set of medical computer-aided diagnosis(CAD)systems,thus significantly improving the accuracy and reliability of medical diagnosis.It has a high practical application in the accurate segmentation of Region of Interest(ROI)in Computed Tomography(CT)for diagnosis,treatment and preoperative planning of diseases.However,medical images are different from images in natural scenes,which often have complex textures and are limited by imaging technology and imaging equipment,and medical images are noisy,blurred and not easy to judge.Therefore,accurate segmentation of CT-based medical images is very challenging and cannot simultaneously combine high accuracy and full automation.To solve these problems,this thesis uses CT medical images of liver and lung in 3D form as the object of study,and uses deep learning methods and traditional segmentation methods to segment the region of interest automatically and with high accuracy,as follows:(1)A set of segmentation and recognition method VLSM-Net model for medical images is proposed,including four steps of image preprocessing,ROI coarse segmentation,ROI fine segmentation,and recognition and classification of ROI.Firstly,in order to avoid the initial bias of the deep learning training model,the CT image is preprocessed,including cropping and resampling;then the preprocessed CT image is used as the input of the improved V-Net model for training,and the training result is used as the initial segmentation result;then the image boundary after the initial segmentation is used as the input,the embedding function of the hybrid level set method is initialized,and the embedding function is gradually updated to complete Finally,the Densenet model is trained to classify and recognize the segmented ROI.(2)It is verified that the hybrid segmentation method combining the deep learning method and the traditional segmentation method can improve the accuracy and precision of segmentation.After the initial segmentation of ROI using the deep learning method,the hybrid level set method(LSM)is further used to refine the segmentation boundary.(3)Experiments were designed to investigate the performance of this hybrid model.In the experiments conducted on the LUNA and Li TS datasets,the VLSM-Net model achieves 92.27% and 82.73% of Dice segmentation scores,respectively.Meanwhile,the segmentation results of VLSM-Net are more refined compared with V-Net,and the model can segment ROI more accurately compared with LSM.The experiments not only demonstrate the effectiveness of the VLSM-Net model in 3D CT image segmentation,but also show that it is feasible to add LSM to the V-Net model to improve the performance,and has some of generalizability.