一种两阶段的由粗到细的多模态脑肿瘤分割框架

A Two-Stage Coarse-to-Fine Brain Tumor Segmentation Framework

从核磁共振图像中提取脑肿瘤在临床诊断及手术规划中起到关键作用。该文提出了一种两阶段的由粗到细的自动分割框架对多模态脑肿瘤图像进行分割。该框架分为粗分割及细分割两部分。粗分割部分采用一个深度卷积神经网络对脑肿瘤五分类,生成4种肿瘤组织的粗分割概率图;细分割部分将这些概率图作为掩膜促使卷积网络关注高概率区域。此外,为了减轻数据不均衡,细分割部分采用了双分支输出,一个支输出五分类结果,并采用带掩膜的交叉熵损失函数;另外一个分支输出二分类结果来标记整个脑肿瘤,采用了均方误差。利用BRATS 2015数据集进行验证,结果表明该方法具有很好的效果。

The accurate extraction of brain tumor from magnetic resonance imaging(MRI)images is the key of clinical diagnostics and treatment planning.A two-stage coarse-to-fine brain tumor segmentation framework is proposed for brain tumor segmentation in multi-modal MRI images.There are two parts in our framework.One is coarse segmentation part,the other one is fine segmentation part.A five-classification task is performed in coarse segmentation part with a deep convolutional neural network.Four coarse probability maps are generated according the five-classification results.The fine segmentation part takes these coarse maps as the mask to guide the network to pay more attention on regions of high probability.Besides,to alleviate the data imbalance problem,there is a two-branch output structure in fine segmentation part.One branch outputs five-classification results with a mask soft-max cross entropy loss function.The other branch outputs a binary result which labels the whole tumor with a mean-square loss function.Our proposal was validated in the BRATS 2015 dataset.It can be proven that our approach achieves a competitive result.