Research On Segmentation Algorithm Of Abnormal Brain MR Image Based On Deep Learning

The abnormal growth of cells in the brain is called brain tumor,which seriously endangers human health.Early detection and timely targeted treatment are crucial for patients.Magnetic resonance imaging(MRI)is the most commonly used auxiliary tool for the diagnosis of brain tumors,and is a mature non-invasive diagnostic medical imaging technology.However,the morphology and internal structure of tumor are diverse and complex,and the boundary between normal tissue and tumor tissue is blurred,which is difficult to distinguish.Manual segmentation is time-consuming and cumbersome,and it is easy to make mistakes.Therefore,based on the deep learning algorithm,this paper proposes a fully automatic and high-precision segmentation method for abnormal brain MR images to assist doctors in diagnosis and treatment.The specific research work of this paper is as follows:(1)Aiming at the problem of information loss and insufficient restoration of important feature information in network deepening,an abnormal brain MRI segmentation method based on dual attention mechanisms is proposed.In order to obtain additional feature information,a convolutional neural network with dual encoders is designed,and the texture feature image generated by T1 image and gray level co-occurrence matrix is input.The residual module is integrated into the network to prevent the gradient from disappearing,and the dual attention mechanisms module is integrated to increase the information extraction of the feature area and reduce the impact of irrelevant information.The Dice coefficients of tumor part,cerebrospinal fluid,gray matter and white matter in the Bra Ts2020 dataset are0.809,0.914,0.909 and 0.925,respectively.(2)The proposed dual attention mechanisms model obtained good segmentation results,but there was a small area of mis-segmentation.Therefore,an improved method based on(1)is proposed to segment the abnormal brain with dual U-Net MRI by fusing multi-scale attention mechanism.The model is composed of two subnetworks with the same structure.DU-sub1 retains the structure of dual encoders,combines the residual module and incorporates the multi-scale attention mechanism in the decoder for feature restoration,making full use of the feature information at different scales.Then it is input to DU-sub2 by multiplying the output of DU-sub1 and T1 image by element,which plays the role of coarse segmentation and fine segmentation respectively.The training was carried out by combining generalized dice loss and classified cross entropy loss.The Dice coefficients of tumor part,cerebrospinal fluid,gray matter and white matter were 0.831,0.917,0.905 and 0.911 respectively.(3)To solve the problem of poor segmentation accuracy of brain tumors in abnormal regions,a multi-scale brain tumor segmentation method is proposed.First,the data is preprocessed,and the brain images of different modes are fused together to complement each other.Then the multi-scale feature fusion module is integrated into the network,and the context feature information is fully obtained according to multiple receptive fields of different sizes.Finally,the residual structure is incorporated to avoid the gradient disappearing.Experiments have proved that this method effectively improves the segmentation accuracy.The Dice coefficients of the whole tumor,core tumor and enhanced tumor on the Bra Ts2020 dataset are 0.876,0.862 and 0.842,respectively.