Research On Dermatological Medical Image Recognition Based On Convolutional Block Attention Capsule Network

Traditional medical image classification methods often design feature extraction methods for specific images,resulting in poor generalization ability and difficult to apply well to other medical images.There are many types of skin diseases,and the similarity between different species is very high,and the differences within the same species are also very large,especially the identification of pigmented skin diseases is more difficult.AI-assisted diagnosis of skin diseases can not only reduce the work intensity of doctors,but also improve the accuracy of skin disease diagnosis.At present,the dermatological recognition method represented by convolutional neural network has achieved good results,but convolutional neural network also has certain limitations,such as the scalar transmission of the upper layer of neurons to the next layer of neurons,and it is difficult to represent the pose relationship between the underlying features and the high-level features.The capsule network is composed of capsules,capsules are a group of neurons composed of vectors,its length represents the probability of the existence of the object,its direction records the object’s attitude parameters(such as position,rotation,etc.),each vector element in the capsule represents a certain attribute(such as posture,position,size,etc.),so the expression ability of the capsule is much stronger than that of a single neuron.Attention mechanisms are better able to focus on key characteristics of the data and ignore irrelevant information.The integration of attention mechanism into the capsule network to recognize dermatological medical images can improve the recognition accuracy.The main work of this article is as follows:(1)The attention mechanism can be used to focus on the advantages of key features and integrate them in the AlexNet network,which can improve the recognition accuracy.The experimental results showed that on the ISIC2018 dermatology dataset,the accuracy of AlexNet integrated into the attention mechanism was improved by 1.9% compared with AlexNet without attention.(2)In convolutional neural networks,scalars are passed between neurons,and it is difficult to represent the pose relationship between the underlying features and the high-level features.Capsules in capsule networks are vectors composed of a group of neurons,which can well represent the position-pose relationship,and a convolutional block attention capsule network model is proposed,which can further improve the accuracy of recognition.The experimental results show that on the ISIC2018 dermatology dataset,the capsule network integrated with convolutional block attention is 4% more accurate than the AlexNet network.(3)In order to solve the problems of excessive reconstruction parameters,slow gradient update and long training time for large-size images of capsule networks,the method of scaling reconstruction is adopted to reduce the number of reconstructed neurons,and improve the training speed while avoiding overfitting.Experimental results show that on the ISIC2018 dermatology dataset,shrinking the reconstructed image can effectively accelerate the training speed.