Research On Medical Image Classification And Segmentation Based On Deep Learning

Although the promotion of medical imaging technology has brought convenience to medical treatment,it has increased the workload of reading images for doctors.Doctors misjudge diseases due to fatigue or lack of professional knowledge,which has a serious negative impact on patients’ physical and mental health and life.Therefore,computer aided medical image diagnosis has become the demand of times.With the development of deep learning technology,medical image classification and segmentation based on deep learning has become a hot topic.We utilize the deep learning technology to focus on tasks with practical clinical values,such as knee osteoarthritis(OA)classification,COVID-19lesions segmentation,etc.The key technologies of medical image classification and seg-mentation based on deep learning are studied to improve the accuracy of medical image classification,segmentation performance and training efficiency of corresponding mod-els.The main work and innovations are summarized as follows:(1)A two-stage knee osteoarthritis classification method in X-rays based on fea-ture enhancement is proposed.Knee osteoarthritis(OA)is a common joint disease that seriously affects people’s daily life.X-ray images are widely used in knee OA diagnosis because of their fast imaging and low cost.At present,most knee OA classification meth-ods are traditional machine learning and shallow convolutional neural network models,which are inadequate in feature extraction and ignore the different importance of global semantic information and local features.To solve the above problems,we design a cas-caded multi-task convolutional neural network to locate knee joint regions more accu-rately.We propose a classification model based on local and global feature fusion and joint loss function,which uses a deeper backbone network and global and local attention,extracting global semantic information and local key features.Experiments on public datasets show that the proposed model enhances feature representation and achieves bet-ter classification accuracy.(2)A fine-grained knee osteoarthritis classification method in X-rays based on Siamese network is proposed.In addition to providing comprehensive knee OA grades,fine-grained knee OA classification can provide additional auxiliary information for doc-tors,so it is also worth studying.Traditional methods do not consider the symmetry of the knee joint region and comparative features,resulting in a lack of semantic information.To solve the above problems,we propose a fine-grained knee OA classification method based on the Siamese network with adaptive gate attention mechanism.In this method,each knee region is divided into two image blocks according to its symmetry and input into the Siamese SE-ResNext50-32x4d network with shared weights to extract local con-trast features.In addition,an adaptive gate attention mechanism module is embedded in the model to enhance the learning of comparative features,and the binary task of knee OA(KL ≥ 2)/non-knee OA(KL≤1)is added to further facilitate feature extraction.Because labels of various knee OA grades are semi-quantitative and fuzzy,we propose a new eval-uation metric: top±1 accuracy,which increases the diversity of evaluation mechanism.Experiments on public datasets show that the proposed fine-grained classification model achieves better classification accuracy.(3)A COVID-19 segmentation method in CT images based on edge detection en-hancement is proposed.Medical image segmentation technology is to segment lesions,helping doctors better evaluate the severity of the disease.For example,COVID-19 le-sions segmentation plays an important role in COVID-19 diagnosis in widely used CT images.Because COVID-19 CT images have blurred lesions boundaries,noise interfer-ence,and lesions edges with diverse shapes,sizes,and locations,the existing COVID-19segmentation methods based on edge detection are insufficient in edge feature extraction.To solve the above problems,we propose an edge detection model called COVID Edge-Net.In the model,the dynamic feature fusion model is used as the backbone network,and multi-scale residual dual-attention module is embedded in the model to enhance edge extraction with different shapes,sizes and positions.Then,the traditional Canny operator is used to extract Canny features to enhance edge sharpening and continuity.Finally,the enhanced edge features combined with traditional features and deep learning features are used in the segmentation task.Experiments on public datasets show that the proposed method can detect more accurate,more consistent,and sharper edges,and improve the performance of COVID-19 lesions segmentation in CT images.(4)A deep learning model optimization method based on learning rate opti-mization strategy is proposed.Medical image classification and segmentation models based on deep learning have a long training cycle and need to tune parameters for many times.Therefore,it is worth studying to improve the training efficiency of deep learning models.At present,learning rate optimization strategies used in training deep learning models have the following issues: they need more manual intervention;learning rates are easy to decay to small values,resulting in slow training convergence;poor generalization performance and high computational complexity exist.To solve the above problems,we present an automatic learning rate decay method based on stochastic gradient descent se-ries optimization algorithms from theoretical inspiration.That is,the learning rate of the k-th step can be expressed by historical learning rates.In the method,only one param-eter is initialized.Gradually decayed learning rates are generated automatically during training,which reduces manual intervention and ensures fast convergence rate and good generalization.Experiments on several public datasets and deep learning models show that the proposed method makes medical image classification and segmentation models based on deep learning more efficient,faster convergence and higher performance.