Glioma Lesion Segmentation Based On AI UNet

Glioma is the most common malignant tumor in the brain,and magnetic resonance imaging(MRI)is a critically important assessment technique for brain tumor diagnosis.Based on characteristics such as location,morphology,signal,peritumoral edema,and enhancement,MRI could assist in the diagnosis of gliomas and prediction of the grade,which is important for further clinical determination of treatment methods and evaluation of prognosis.Recently,the application of artificial intelligence in medical imaging diagnosis is also in full progress.Radiomics and deep learning have shown great application prospects in the classification and detection of gliomas.Furthermore,the segmentation of glioma plays an important role for clinicians.Traditional manual segmentation made by neurosurgeons and radiologists is time-consuming and cumbersome,so the innovation of automatic and efficient segmentation methods is extremely urgent.This thesis proposed an Attention based Inception U-Net(AI UNet)to give a more efficient segmentation method for glioma,where a new block—Attention based Inception Block(AI Block),combining convolution and Self-Attention,is introduced.This module combines the smaller receptive field of convolution with the larger receptive field of Self-Attention,so as to extract more diverse feature maps to meet the needs of refined segmentation.Several AI Block deformation will be introduced and applied to the segmentation of glioma lesion,with the U-shaped network fusion constructs the AI UNet image segmentation,this network has excellent segmentation performance,as well as low amount of parameters and calculation.Moreover,a loss function combined with GHM loss and Dice loss is plugged into the network,thereby improving the robustness of the network.Experiments show that the proposed network and loss function can improve segmentation effects comparing with state-of-the-art methods.