Room: Room 207
Purpose: Glioblastomas (GBM) are often highly aggressive malignancies with poor prognosis due to its clinical and molecular heterogeneity. Accurate segmentation of the GBM tumor plays a key role in GBM diagnosis and radiotherapy. The purpose of this work is to develop an automatic learning-based segmentation method for GBM tumors in multiparametric MRI (mp-MRI).
Methods: A patch-based fully convolution network (FCN) model was proposed to classify the tumor or non-tumor voxels in mp-MRI (T1W, T2W and FLAIR). Our 3D FCN model adopted the architecture of V-Net, in which the cross-layer connections were used to preserve the resolution of the original input. To overcome the class imbalance problem of tumor in MRI, we sampled a balanced training patch set and then adopted a two-stage training strategy. In the first stage, we enabled batch normalization to learn the empirical normalization parameters for the input batch. In the second stage, we kept the batch normalization parameters as constants and trained the convolution layer parameters. Moreover, in order to enlarge the perceptive field of the neural network to capture more global spatial information, we introduced a multiscale strategy to further refine the segmentation performance. Finally, the well-trained FCN model adaptively labeled GBM tumor in mp-MRI.
Results: This learning-based segmentation technique was tested using the BRATS 2015 dataset (200 patients for the training and 74 patients for our validation). The accuracy of our approach was assessed using the manual segmentation. With the two-stage training strategy and the multiscale smoothing method, the mean volume Dice similarity coefficient was 0.87Â±0.06 between our and manual segmentation.
Conclusion: We developed a brain tumor segmentation approach based on a patch-wise multiscale V-Net model for mp-MRI, demonstrated its clinical feasibility, and validated its accuracy with ground truth. This segmentation technique could be a useful tool for image-guided interventions in brain tumor.
Not Applicable / None Entered.