Application Of Convolutional Neural Network And Attention Mechanism In Medical Image Segmentation And Recognition

Currently,clinicians to the diagnosis of many diseases are relying on the observation and analysis of medical images.However,due to the limitation of imaging quality of imaging equipment and the limitation and influence of clinicians’ experience and knowledge,doctors often make misdiagnosis and missed diagnosis.To alleviate this problem,computer-aided diagnosis came into being.In recent years,convolutional neural network has been introduced into various applications in the field of medical image,such as disease classification,prediction,lesion detection and image registration,and has achieved great success.Compared with the traditional machine learning methods,based on the depth of the convolutional neural network learning method can characteristics of the engineering of multifarious,convolution neural network powerful expression ability of features,an end-to-end complete the task,and in recent years,there are some studies to explore the use of attention mechanism in convolution neural network to improve the network performance.However,the application of convolutional neural network in medical image still faces many challenges,including insufficient model training and poor performance due to the small amount of medical image data.The imaging quality of some imaging instruments is poor,which increases the difficulty of model optimization.In view of the challenges of convolutional neural network in the field of medical image,the application of convolutional neural network and attention mechanism in the field of medical image was discussed in this paper with the objects of Automated Whole Breast Ultrasound(ABUS)and fundus retinal image.The main research results include the following two aspects:On the one hand,for the segmentation of breast anatomy for ABUS image,this paper proposes a method of segmentation of breast anatomy for ABUS image based on co-attention.With the Res Ne Xt as the infrastructure,the spatial and channel attention modules are embedded as the coding path,and the non-local context block is further introduced to capture the long-range dependencies by calculating the relationship between any two locations in the feature map,thus helping to improve the segmentation performance.The decoding path of this method adopts the weighted up-sample block in order to get a better sampling effect on the class specificity.At the same time,the co-attention mechanism is introduced to construct the network into a double-input network,so as to learn the correlation between the continuous slices,so as to improve the segmentation consistency between the continuous slices.Experiments have verified the effectiveness of this method for the anatomical segmentation of complex ABUS images.On the other hand,for the recognition of retinopathy of prematurity,this paper uses attention residual network to classify.Specifically,we first selected Res Net as our infrastructure and embedded the spatial and channel attention modules to enhance its feature representation.Then,we used the gradient-weighted class activation mapping(Grad-CAM)to visualize the trained model and explore the interpretability of the network.The effectiveness of the method was verified by experiments,and the model also successfully detected the lesion structure(dividing line or ridge line)of ROP in the retinal image.To sum up,this paper discusses the application of convolutional neural network in medical image segmentation and recognition based on Automated Whole Breast Ultrasound Image and fundus retinal image,and uses the attention mechanism to enhance the feature expression ability of the network,which is realized by focusing on useful features and suppressing unimportant features.In addition,co-attention can also explore the correlation between different inputs.In this paper,it is used to calculate the correlation of AUBS continuous slices to maintain the consistency of continuous slice segmentation.In this paper,a large number of comparative experiments are carried out to verify the effectiveness of the proposed method.This method is not only limited to automated whole breast ultrasound image and fundus retinal image,but also can be extended to other medical image analysis.