多模态MR图像和多特征融合的胶质母细胞瘤自动分割

胶质母细胞瘤(glioblastoma multiforme, GBM)是恶性度最高的脑胶质瘤,其病变组织的定位和量化计算对肿瘤的诊断及制定治疗计划至关重要.为提高GBM自动分割的准确性,提出一种多模态MR图像和多特征融合的GBM自动分割算法.首先在图像配准和偏置场校正后,融合GBM多模态MR图像提取各体素的多个底层特征,构建随机森林(random forest, RF)模型,依据特征信息粗分割;其次将多种子点三维区域生长分割GBM多模态MR图像的结果替换相应置信度低的粗分割结果,生成训练数据重新训练RF模型,精分割GBM多模态MR图像;最后考虑GBM解剖结构先验知识、阈值分割和中值滤波精分割结果后得到最终结果.以平均Dice相似性系数、Hausdorff距离和敏感度为评价指标,该算法分割GBM-nih-zcmu数据库中整个肿瘤的平均Dice相似性系数、Hausdorff距离和敏感度分别为0.879, 6.232和0.863,能有效地提高GBM多模态MR图像自动分割的精度,满足临床应用对准确率的要求.

Automatic Segmentation for Glioblastoma Multiforme Using Multimodal MR Images and Multiple Features

Glioblastoma multiforme(GBM)is the most malignant glioma,and localization and quantification of diseased tissues is crucial for tumor diagnosis and treatment planning.To improve the accuracy of GBM automatic segmentation,a novel GBM automatic segmentation method is proposed based on multimodal MR images and multiple features.Firstly,after image registration and bias field correction,multiple low-level features of each voxel were extracted from the GBM multimodal magnetic resonance(MR)images,and a random forest(RF)model was constructed for coarse segmentation according to the features information.Secondly,the corresponding coarse segmentation results with low confidence were replaced by the results of multiple seeds three dimensional region growing segmentation for GBM multimodal MR images,and the RF model was retrained using the generated training data,then fine segmentation of GBM multimodal MR images was implemented.Finally,considering the prior knowledge of the GBM tumors anatomical structures,the final results were achieved after threshold segmentation and median filtering of the fine segmentation results.The proposed algorithm adopts the average Dice similarity coefficient,Hausdorff distance and sensitivity as the evaluation index,the average Dice similarity coefficient,Hausdorff distance and sensitivity of the entire tumor in the GBM-nih-zcmu database are 0.879,6.232 and 0.863,respectively,which effectively improve the accuracy of automatic segmentation of GBM multi-modal MR images and meet the accuracy requirements of clinical applications.