Research On Brain Tumor Image Classification Technology Based On Deep Learning

Brain tumor is a brain disease.Early detection and accurate judgment of tumor type and lesion degree can gain precious treatment time for patients.However,the diversity of brain tumor shape,the unfixed location of the disease and the difference of the lesion degree in different patients make the accurate classification of brain tumors extremely challenging.At present,manual diagnosis is usually used in clinical practice to identify tumor types and lesions.Although the method of manual diagnosis has high accuracy,in the face of massive image data,doctors will inevitably misdiagnose,and manual diagnosis is time-consuming.Therefore,the research of brain tumor image automatic classification has important application value.In recent years,deep learning has been widely used in various fields.Taking medical images as an example,the image classification method based on deep learning can autonomously learn the deep information in the data,which can speed up the diagnosis and improve the accuracy of the diagnosis.Misdiagnosis caused by human factors can be avoided.In this paper,we carried out research on the classification technology of brain tumor lesions and tumor types,and completed the following work:(1)A classification model was proposed to classify benign and malignant brain gliomas which combines multi-scale feature extraction module based on dilated convolution and channel attention mechanism.The model used Res Ne Xt network,a variant of the deep residual network,as the backbone network,and introduced a multi-scale feature extraction module based on dilated convolution to replace the first convolution layer of the network in order to expand the receptive field while retaining the image resolution and fuse the global feature with the subtle feature.Then the module of channel attention mechanism was introduced to integrate the information of the feature channel and improve the weight of the tumor area.This classification model solved the problems such as low accuracy of benign and malignant classification,and mostly non-end-to-end classification systems.The accuracy of the proposed model in this paper was improved by 2.85 % and 0.65 %compared with Res Ne Xt in the Bra TS2017 and Bra TS2019 datasets,respectively.(2)A classification model was proposed to classify glioma,meningioma and pituitary tumor which was based on Squeeze and Excitation module and spatial attention mechanism.The Res Ne Xt network is selected as the backbone network,and the improved Squeeze and Excitation module and spatial attention mechanism module were introduced into the Res Ne Xt structure to integrate the channel information and spatial location information of key features,so as to improve the attention of the network to key features.This classification model solved the difficulties caused by the difference of the location and imaging features among different types of tumors,the shape diversity and the non-fixed location of tumors in the same category.Compared with Res Ne Xt,the recall rate of the proposed model in this paper was improved by 1.25 %,5.25 % and 1.40 % for glioma,meningioma and pituitary tumor in the CE-MRI dataset,respectively.(3)In the process of training the network,an optimization strategy combination was proposed to optimize the learning rate,image label value and pre-training of transfer learning.It solved the problems of complex network training process and overfitting.The optimization strategy combination was applied to the two classification tasks in this paper.The improved MDCA-Res Ne Xt network achieved 98.11 % and 98.72 % accuracy in the Bra TS2017 and Bra TS2019 datasets,respectively.The improved Res Ne Xt-SESA network achieved 98.69 % accuracy in the CE-MRI datasets.