Adaptive Window for Automatic Classification-Based Segmentation of Multimodal Brain Tumor

Classification-based segmentation methods for multimodal brain tumor segmentation suffer from a serious problem. This problem comes from computing features around voxels in the MRI volume using a fixed window of predefined size and shape. In this case for voxels at or near boundaries of tissues, the fixed window will contain voxels belonging to more than one class, thus yielding to erroneous feature calculations and voxel misclassifications. To solve this problem, the paper proposes a novel method of using an adaptive window where the window's size and shape are computed based on the MRI volume around the voxel in concern. The paper compares the performance of the proposed adaptive window and those of fixed window-based methods in the application of automatic brain tumor segmentation. The dataset used in the experiments is 3D MRI volumes from MICCAI 2016 brain tumor segmentation (BRATS) benchmark. Our experimental results indeed demonstrate the improvement in performance in terms of the Dice metric with the proposed method.