多分类CNN的胶质母细胞瘤多模态MR图像分割

为提高胶质母细胞瘤（GBM）多模态磁共振（MR）图像中各肿瘤子区域分割的准确性，提出一种多分类卷积神经网络（CNN）的GBM多模态MR图像自动分割算法.首先在98%缩尾处理和配准GBM多模态MR图像后，利用N4ITK法校正偏移场；其次构建一个主要由4个卷积层、2个池化层和2个全连接层组成的多分类CNN模型，训练后预分割GBM多模态MR图像，将体素分为5类不同的标签；最后移除所有小于200体素的假阳性区域，中值滤波后获得最终分割结果.以Dice相似性系数DSC、阳性预测值PPV和平均Hausdorff距离AHD为评价指标，利用所提出的算法对F-C-GBM数据集中整个肿瘤组织进行分割，获得的DSC、PPV、AHD分别为0.889±0.087、0.859±0.127和1.923.结果表明，该算法能有效提高GBM多模态MR图像分割的性能，可望有临床应用前景.

Glioblastoma Multiforme Multi-modal MR Images Segmentation Using Multi-class CNN

To improve the accuracy of segmenting the tumor sub-regions in glioblastoma multiforme (GBM) multi-modal magnetic resonance (MR) images,a GBM multi-modal MR images automatic segmentation algorithm is proposed by using multi-class convolution neural network (CNN).Firstly,after 98% winsorization and registration for the GBM multi-modal MR images,the bias field was corrected by using the N4ITK method.Secondly,a multi-class CNN model mainly consisting of four convolutional layers,two pooling layers and two fully connected layers was constructed;the GBM multi-modal MR images were pre-segmented after training,and voxels were classified into five different labels.Finally,all false positive regions smaller than 200 voxels were removed,and the final segmentation results were obtained by median filtering.The Dice similarity coefficient DSC,positive predictive value PPV and average Hausdorff distance AHD were adopted as the evaluation index,and the DSC,PPV as well as AHD were 0.889±0.087,0.859±0.127 and 1.923 for segmenting the entire tumor tissues in F-C-GBM dataset by the proposed algorithm,respectively.Results indicate that the proposed method can effectively improve the performance in the segmentation of the GBM multi-modal MR images and may be expected to have clinical application prospects.