| According to the statics from National Cancer Center,brain cancer is the second most common childhood cancer and the third most common cause of cancer death in adults aged 15-34.The cure rate of brain cancer has not increased with the advancement of technology.Only be detected and treated at an early stage can effectively improve the cure rate and prolong survival time of patients.With the advancement and promotion of medical imaging technology,MRI has become the main diagnostic method for brain cancer.It can help doctors diagnose brain cancer without surgery.However,the interpretation of medical images still requires doctors to have extensive clinical experience,and screening detailed brain images takes a lot of time and energy.These are unfavorable factors for the diagnosis and treatment of brain cancer.With the continuous progress and development in the field of machine learning in recent years,the use of computer vision technology to assist doctors in interpreting medical images has become a hot research direction and was tested with COVID-19.Previous machine learning methods have been designed by researchers to extract features by themselves.It has high requirements on the designer’s medical and computer skills,which greatly increases the development cycle and labor costs,and is not conducive to the deployment and adjustment of the program.Due to the rich details of the brain image,the image scaling,cropping and other modifications will affect the accuracy of the judgment result.At the same time,the overall incidence of brain cancer is not high.Because of medical ethics and other reasons,there are fewer samples that can be used for learning,and the distribution of various categories is unbalanced.These are difficulties and challenges to the diagnosis and classification of brain cancer.In response to the above-mentioned difficulties and challenges,this paper proposes a new type of non-pooling Res Capsule network with scaled reconstruction model.The main work of this paper is as follows:1、Using non-pooling convolution structure with shortcut to extract high-dimensional features.For medical images with a large size,the previous methods generally reduce the size of the input data by cropping,pooling,etc.But the above method also relies on experience and takes time.We use non-pooling convolution structure with shortcut to extract high-dimensional features.In order to prevent the relative position relationship of features from being damaged,we removed the pooling layer;at the same time,while avoiding network degradation,we increased the depth of the network,increased the data dimension,and reduced the amount of data.2、Using dynamic routing agreement to achieve better results on small data volumes and unbalanced datasets.Compared with the fully connected layer that does not consider the relative position of the data and the combination relationship,we divide the data into multiple capsules and calculate the final capsules through dynamic routing agreement.The agreement effectively learns and utilizes the translation invariance and rotation invariance of the data,and improves the performance of the model on small datasets and unbalanced datasets.3、Using scaled reconstruction to increase calculation speed while avoiding overfitting.The final capsule has the information of the category it represents,and the original input can be roughly restored by reconstruction.Adding a certain weight of reconstruction loss to the loss can avoid overfitting.However,in large-size images,full reconstruction of the image requires a lot of computing resources,and gradient vanishing and gradient explosion are prone to occur.We calculate the reconstruction loss with the scaled picture,which improves the calculation speed while ensuring the model’s antioverfitting ability.4、Based on the public dataset from kaggle,the validity of the proposed classification model is verified,and test the performance of the model under different amount of data. |
PDF Full Text Request