Research On Glioma Image Segmentation Method Based On Multi-sequence MR Image Fusion

Glioma is a disease that seriously threatens human health.The fine segmentation of glioma images can help doctors find cancer cells more effectively and help in future diagnosis.Due to the invasive growth of cancer cells,there is no clear boundary between them and normal brain tissue,resulting in a lot of challenges in the task of segmentation.At present,the segmentation of gliomas mainly relies on doctors to perform manual segmentation,which will not only increase the burden on doctors and take up a lot of manpower,but also require the segmentation personnel to have the necessary medical knowledge,Therefore,more and more researchers have begun to pay attention to automatic segmentation of glioma images based on deep learning.In recent years,in many image segmentation competitions,people rely on deep learning to achieve a series of results.This thesis has done the following work for glioma image segmentation:(1)Taking the U-shaped convolutional neural network(UNet)as the basic structure,considering the insufficient information acquisition of the network,a two-dimensional UNet segmentation network model based on the dilated convolutional pyramid module is proposed.By introducing the dilated convolution pyramid,the receptive field of the convolution kernel is expanded,so that the model can obtain multi-scale features.Correction and registration of data to enable images to filter noise and maintain optimal informative features.At the same time,considering the small size of the dataset,data augmentation techniques are used to expand the dataset.(2)In order to make full use of the spatial information between different slices,this thesis selects the 3DUNet network as the basic structure,and proposes a 3D-level improved UNet for glioma image segmentation,and the convolution module in the UNet network is improved into Residual module,which simplifies network training while ensuring skip connections in residual units and information propagation between low-level and high-level networks without gradient disappearance.Finally,the squeeze-and-excitation network(SE-Net)is introduced in the cascaded part of the network,and it is used as an attention mechanism module,so that the network can pay more attention to the key regions.(3)A 3D glioma image segmentation method based on multimodal fusion is proposed.The glioma images contain four sequences,and each sequence has different characteristics.In order to make full use of the information of the four sequences,this thesis adds a multi-sequence fusion module before data input.At the same time,in order to reduce the complexity of the network,an efficient channel attention network(ECA-Net)is introduced in the cascade part.And introduce the dilated convolution pyramid module in the decoding area to expand the receptive field.The three work together to improve the performance of the network for glioma image segmentation.