胶质母细胞瘤多模态磁共振图像自动分割

针对大多数传统胶质母细胞瘤（GBM）多模态磁共振（MR）图像分割算法未能将整个肿瘤细分为不同子区域的问题，提出基于混合特征和先验知识的GBM多模态MR图像自动分割算法. 配准GBM多模态MR图像，将头部区域方位调整到近似未旋转位置，并利用N4ITK法进行偏置场校正. 在提取GBM多模态MR图像局部位置特征、强度特征、纹理特征、对称特征和上下文特征后，应用随机森林分类器初步分割GBM多模态MR图像. 考虑GBM肿瘤解剖结构先验知识，移除小区域和中值滤波后得到最终分割结果. 以Dice相似性系数作为评价指标，利用所提出的算法对TCGA-GBM和CH-GBM数据库中整个肿瘤进行分割，获得的平均Dice相似性系数分别为0.871、0.882. 结果表明，该算法能以较高的准确率分割GBM多模态MR图像，适用于临床GBM多模态MR图像分割任务.

Automated segmentation for multi-modal magnetic resonance image of glioblastoma multiforme

A glioblastoma multiforme (GBM) multi-modal magnetic resonance (MR) image automated segmentation algorithm based on hybrid features and prior knowledge was proposed, as most traditional GBM multi-modal MR image segmentation algorithms failed to subdivide the whole tumor into different sub-regions. The head region was adjusted to the approximate unrotated position once the GBM multi-modal MR image was registered, and the bias field correction was performed by the N4ITK method. A random forest classifier was applied to initially segment GBM multi-modal MR image after the extraction of the local location features, intensity features, texture features, symmetric features and contextual features of GBM multi-modal MR image. The final segmentation results were obtained by removing small regions and median filtering, based on the prior knowledge of the anatomical structure of GBM tumor. The Dice similarity coefficient was adopted as an evaluation metric, and the average Dice similarity coefficient values were 0.871 and 0.882 for segmenting the whole tumor in TCGA-GBM and CH-GBM databases by the proposed algorithm, respectively. Results indicated that the proposed method is suitable for clinical application of GBM multi-modal MR image segmentation task with relative high accuracy.