Research And Implementation Of Brain Tumor Image Classification Based On Capsule Network

Brain tumors are one of the top ten malignant tumors with high morbidity and mortality.Early and accurate diagnosis can help doctors propose treatment plans for patients and reduce mortality.Using artificial intelligence to design a disease classification system can optimize the diagnosis results in the examination.At present,the classification of brain tumors faces the problem of fewer samples in the data set.Convolutional neural networks cannot use the spatial position relationship between the features in the image,and their performance on the task of medical image classification is insufficient.To this end,this thesis has designed and implemented the following work content.1.The attention capsule mechanism is added to the capsule network.An attention capsule layer is added between the convolutional layer and the capsule layer.The iterative attention algorithm is used in the capsule attention layer to make the network have dynamic capsule attention,and the output capsule with prediction information can help increase the weight ratio of key features.It solves the problem that other brain tissue information contained in MRI images of brain tumors interferes with the training process.The network model with the attention capsule layer added in the experiment can effectively improve the accuracy of target classification.2.Improved the connection algorithm of capsule routing between successive capsule layers.The connection mode of the capsules between successive capsule layers uses the central route.The central route uses the similarity measurement for the input feature vector,and adds the measurement to all the prediction networks as the input of the highlevel capsule.All the capsule types in the continuous capsule layer are cyclically traversed to realize the update and propagation of the capsule types.In experimental verification,the central routing solves the problem that the capsule feature pose matrix in the highlevel capsule layer involves many parameters,and can effectively improve the classification accuracy of three brain tumors in the brain magnetic imaging of meningiomas,gliomas and pituitary tumors.3.Optimized the convergence speed of brain tumor classification model training.In the loss function design of the Softmax layer,the weight coefficient is added to the crossentropy loss function,and the loss function in the capsule layer is combined with the central route to add the calculation of the capsule metric.The optimized loss function reduces the training time of the network model.In the experimental stage,the average number of iterations to reach the local optimal solution in the five-fold cross-validation of the model is reduced from the original 20 times to 16 times.4.A classification system for brain tumors has been implemented.The classification system includes a brain tumor detection module and a case management module.The case management module includes the realization of the interface and the test analysis of the stability of the interface.The brain tumor classification system includes the abovementioned three work contents.The imaging method provides certain reference opinions in the daily diagnosis work of doctors,and reduces the probability of errors caused by human factors.