A Study On 2D-3D Cascade Network For Glioma Segmentation In Multisequence MRI Images Using Multi-Scale Information

Glioma is the most commonly malignant brain tumor in adults,which shows invasive growth and lacks a clear boundary with normal brain tissue.Magnetic resonance imaging(MRI)is the most commonly used imaging method for glioma.Glioma segmentation plays an important role in guiding the operation,assisting in delineating the radiotherapy target,and evaluating the curative effect and prognosis.However,manual delineation of tumors is time-consuming and laborious,as well as with poor reproducibility and subjective differences.Although image segmentation algorithms based on deep learning can achieve automatic glioma segmentation,it is still a challenging task for the following reasons:first,the sizes,shapes,and locations of gliomas vary greatly in the brain.Second,due to the aggressive growth of gliomas,their edges are complex and blurred.Third,intensity values of MRI images varied greatly due to different machine types and acquisition protocols.Finally,many MRI images consist of anisotropic dimensions with high variations along the z-axis direction in clinic,which further increases the difficulty for automatic segmentation.UNet-based networks are widely used in medical image segmentation tasks and have achieved state-of-the-art performance.However,context information along the z-axis is ignored by using 2D convolutions,whereas computational cost is considerably high in 3D convolutions.Moreover,original UNet structure cannot capture finer details.To address these issues,a novel 2D-3D cascade network with multi-scale information module is proposed in this study for the multiclass segmentation of gliomas in multisequence MRI images.First,a multi-task learningbased 2D network is applied to fully exploit potential intra-slice features.A variational autoencoder module is incorporated into the 2D DenseUNet to regularize the shared encoder to extract useful information and to represent glioma heterogeneity.Second,a 3D DenseUNet is integrated with the 2D network in cascade mode to extract useful inter-slice features.Moreover,a multi-scale information module is used in both 2D and 3D networks to further capture the finer details of gliomas.Finally,the whole 2D-3D cascade network is trained in an end-to-end manner,where the intra-slice and inter-slice features are fused and optimized jointly to take full advantage of 3D image information.Our method is verified on the publicly available BraTS 2018 dataset.For the whole tumor(WT),tumor core(TC),and enhanced tumor(ET),Dice values reach 0.911,0.863,and 0.808,respectively,and Hausdorff distances reach 3.923,5.798,and 2.889,respectively.The proposed method is performed on a clinical data set to further evaluate its generalized ability.Dice values of WT and TC are 0.873 and 0.767,respectively,and the Hausdorff distances are 6.435 and 8.275,respectively.Compared with other advanced methods,our method achieves competitive performance,which shows that our glioma segmentation method is feasible and effective.