Lightweight Deep Network For Brain Tumor Segmentation From Multimodal Magnetic Resonance Images

Brain tumor is one of the most common diseases that cause great harm to the human body,with high morbidity and mortality.The analysis of magnetic resonance images of brain tumors is an important basis for doctors to perform brain tumor diagnosis and treatment,surgical evaluation and disease tracking.However,brain tumors have various shapes and complex structures.Manual segmentation is very time-consuming and labor-intensive,and misjudgment can easily occur during segmentation.Therefore,in order to simplify the segmentation process of brain tumor magnetic resonance images,improve the work efficiency of medical workers,and obtain more accurate segmentation results,a fast and fully automatic brain tumor segmentation algorithm has very important clinical practical value.This paper proposes two effective segmentation methods for multimodal brain tumor segmentation tasks.The main idea is to use the structural characteristics of brain tumors to achieve a competitive segmentation accuracy within a limited computing budget.The specific research work is as follows:(1)For the problems that the current 3D network has large computing memory and low computing efficiency,a lightweight hierarchical decoupling convolution and multi-view fusion network model—HDC-Net~0 is proposed,which completed 3D segmentation of brain tumors based on 2D convolution and achieved excellent segmentation accuracy.The model integrates a view decoupling convolution module,a hierarchical group convolution module and a hierarchical decoupling convolution module,so that the model can capture the 3D MR image spatial context information and multi-scale information through a lightweight structure.In addition,due to its unique design structure,the model can be flexibly integrated with strategies such as multi-view fusion and model cascading to further improve segmentation performance.(2)For the design defects of HDC-Net~0,such as high system complexity and inability to accept large-scale image input,a lightweight and more concise and efficient HDC-Net model was proposed to complete multi-class segmentation of brain tumors at one time and further improve the segmentation accuracy of brain tumors.Specifically,an additional down sampling strategy is introduced to allow the model to accept larger scale image input.In addition,combined with the flexible design structure of the hierarchical decoupling convolution module,HDC-Net model with various improved structures are proposed to further improve the segmentation performance of the model.This paper makes detailed experimental demonstrations of the two proposed segmentation methods on two challenging BraTS 2017 and BraTS 2018 public datasets,respectively.The results show that the method proposed in this paper obtains competitive segmentation results,and the model has a very small amount of parameters,which is very suitable for use with limited equipment resources.