Research On MRI Brain Tumor Segmentation Algorithm Based On Feature Fusion And Attention Mechanism

Glioma is the most common and aggressive primary brain tumor,accounting for about half of all intracranial primary tumors and posing a serious risk to human health.However,the common brain tumor segmentation operation in clinical practice is still manual segmentation,which is very time and energy consuming for physicians,and the segmentation results vary greatly depending on their personal experience.At the same time,brain tumors have different shapes and sizes,blurred borders and unbalanced categories,which make the automatic segmentation of brain tumors more challenging.Therefore,it is important to study an automatic brain tumor segmentation technique.In this paper,we take multimodal brain tumor data as the research object,and use the strategy of feature fusion and attention mechanism to build a fully automated brain tumor segmentation model with high accuracy and efficiency for the specific problems of brain tumor segmentation,as follows.(1)A brain tumor segmentation model based on multiscale feature fusion is proposed to address the problems of varying shapes and sizes of brain tumors as well as class imbalance.Firstly,the proposed model uses the null feature pyramid module to extract and fuse the multi-scale features of brain tumors to segment brain tumors of different shapes and sizes.Secondly,the proposed model uses the depth residual module to reduce the number of parameters and improve the network depth while avoiding network degradation.Then,the proposed model increases the weight of effective features of brain tumor through the channel attention mechanism module,avoiding the redundancy of background and other information.In addition,a mixed loss function is used to alleviate the category imbalance problem of brain tumor segmentation.Finally,the Dice coefficients of ET,WT,and TC of the proposed model on the BraTS2018 validation set reached 0.77,0.88,and0.82,respectively.(2)A brain tumor segmentation model based on a triple efficient attention mechanism is proposed to address the problems of ambiguous brain tumor boundaries and high computational complexity of the model.Firstly,to address the limitations of the dimensionality reduction operation in the channel attention mechanism,the proposed model uses an efficient channel attention mechanism to solve the problem.The module contains a local cross-channel interaction strategy without dimensionality reduction,which effectively avoids the effect of dimensionality reduction on the channel attention learning effect,enhances the feature extraction of tumor edge detail information,and reduces the number of parameters to improve the computational efficiency.Secondly,the proposed model introduces a triple parallelism strategy to capture cross-dimensional interaction information by transposing the input features,which enriches the utilization of spatial information;in addition,the convolution methods used in the proposed model are all deep over-parametric convolution,which improves the segmentation accuracy while accelerating the convergence of the model.Finally,the Dice coefficients of ET,WT,and TC of the proposed model on the BraTS2020 validation set reached 0.80,0.89,and 0.83,respectively.To evaluate the model,three datasets,BraTS2018,BraTS2019,and BraTS2020,were used for experiments and online validation were conducted in this paper.The experimental results show that the brain tumor segmentation model based on multi-scale feature fusion and triple efficient attention-based mechanism has good segmentation performance and is challenging to compare with other studies.