Enhancing Convolutional Features For Multi-modality Brain Tumor Image Classification

Medical image classification is one of the fundamental research topics in the domain of computer-aided diagnosis.This technology provides assisted diagnosis advice to doctors by processing medical images,which increases the accuracy of diagnosis and helps patients avoid painful and time-consuming pathological analysis.This thsis conducts the research on multi-modality brain tumor image classification.First,we consider that the current existing natural image classification models cannot be reliably applied to the classification for brain tumors.This is because of the uncertainty of location,shape,structure for brain tumors and the lack of the MRI samples.Second,the dataset for brain tumors used in this thesis contains three modalities(FLAIR,T1 and T2).Effectively fusing the multi-modality images can makes models obtain more comprehensive visual information.Based on the above two aspects,we propose an image classification model for brain tumors and a multi-modality feature fusion method.The main research contents for this thesis are as follows:1)In this thsis,a brain tumor image classification model based on enhancing convolutional features is proposed.In this model,we design two forms of attention mechanism modules,which add spatial enhancing convolutional features and channel enhancing convolutional features in series and parallel respectively.The attention mechanism modules assign a weight to each neuron of the feature maps,which makes the classification models pay more concentration to the discriminative lesion structures of brain tumors.Besides,the proposed model takes advantage of the small sample learning of Res Net and makes it become the baseline.The designed attention mechanism modules are incorporated in each residual block of Res Net,and then two forms of Discrimination Oriented Network(DONet)for brain tumor image classification are obtained in this thesis.2)In this thesis,a multi-modality feature fusion method applied for medical scenes is proposed.This method connects feature maps from different modalities in channel dimension firstly,and then inputs the obtained feature map into the three–branch structure — spatial branch,channel branch and identity branch.Spatial branch integates the visual information from different modalities and obtains a two–dimensional distribution of weights in spatial direction,helping our classification model achieve better localization of lesion areas.Channel branch fully considers the correlation between different modalities and captures a one–dimensional distribution of weights in channel direction,helping our classification model focus on the modalities with richer visual information and stress the modalities with interference information.Finally,weights learned from spatial and channel branch are multiplied with the feature map of identity branch in this method,obtaining a feature map with multi-modality visual information.