Research On Classification And Segmentation Of Nasopharyngeal Carcinoma And Cardiac MRI Medical Images Based On Deep Learning

Medical imaging techniques such as magnetic resonance imaging(MRI)provide effective anatomical imaging for the segmentation of nasopharyngeal carcinoma(NPC),segmentation and classification of myocardium.In the process of clinical diagnosis,doctors usually need to manually mark the boundaries of NPC or myocardium to assist the treatment of patients.This process is tedious,time-consuming and error-prone.Therefore,the use of computer-based automatic segmentation algorithm to assist doctors to accurately locate NPC and myocardial regions has become a necessary requirement.However,due to the small area of NPC,it is easy to cause extreme imbalances in the foreground and background categories.In addition,the tissue structure(such as shape,size,etc.)of each patient’s cancer area is quite different,and the MRI medical image has special characteristics,which makes the automatic segmentation task of NPC more challenging.However,myocardium segmentation and classification has been a challenging task due to low contrast and large variations in intensity and shapes.In addition,inherent noise in synthetic cardiac magnetic resonance(CMR)images caused by motion artifacts and heart dynamics would interfere with myocardium analysis.Therefore,deep learning based on NPC and CMR medical image classification and segmentation technology has extensive research value and application prospects.The work of this study is the classification and segmentation of NPC and cardiac MRI medical images,which are mainly divided into the following aspects:(1)Due to the problem of poor quality and labeling of medical images,we first research the outlier detection algorithm to process medical images,and to screen poor quality and labeled images.We propose a new distance-based outlier detection method that depends on data structures to detect these outliers.In the proposed method,a global binary tree is constructed and the local distance score(LDS)of a point is calculated to evaluate to what degree the observation is an outlier.The greater the value of the LDS,the more likely the point is an outlier point.Unlike typical distance-based methods,our algorithm has good scalability.Even when the dimension of the data points increases,the performance of our algorithm does not diminish.To reduce extra parameters,the top-p ranked points can be identified as outliers.Experimental results on synthetic and seven UCI real-world datasets demonstrate the effectiveness and stability of our method.(2)For the problem of difficult segmentation of the three-dimensional NPC dataset,we propose a three-dimensional convolutional neural network with multi-scale feature pyramids,which can learn the features of NPC with different sizes and shapes.In this method,we use atrous convolutions with different atrous rates to accommodate different sizes of images in the network.We used Jaccrad as a loss function to handle extreme imbalances in foreground and background category.Experiments show that our method can prevent a large number of negative voxels from moving away from the boundary in the training process,so that the segmentation algorithm can correctly segment the location of NPC.In the experiment,we trained and tested 3D MRI images of 120 clinical patients,and the calculated average Dice similarity coefficient was 0.7298.The results show that our approach exceeds the segmentation results of the other four major network structures and is comparable to the level of experience of physicians.(3)For the problem of difficult segmentation and classification of the two-dimensional myocardial dataset,inspired by the shared representation between related tasks,we propose a unified framework for left ventricular(LV)myocardial segmentation and classification,and at the same time achieve segmentation of myocardial regions and classification of myocardial MRI images,such as normal human myocardium and patient myocardium.In the proposed method,we first establish an effective cascading atrous convolution to fuse features with various receptive fields.By using various atrous rates,we can obtain multi-scale features and hierarchical context information regarding myocardium,which can simultaneously improve the performance of both myocardial segmentation and classification.We evaluated the proposed unified framework on 488 CMR T1-mapping images with lower peak signal-to-noise ratio.The Dice metric for automatic segmentation is 0.8173,and the Accuracy metric of 0.9734 for automatic diagnosis is reported on this data set.The results show that the segmentation results of our unified framework are better than other methods.