基于多个密集连接型2D-CNNs的脑胶质瘤MRI三维分割

Segmentation of brain glioma on MRI using multiple densely connected 2D-CNNs

准确可靠的脑胶质瘤分割是脑胶质瘤诊断、治疗方案制定和治疗效果评价的重要前提。为了有效针对脑胶质瘤MRI的特性和基于CNNs的脑胶质瘤分割方法的不足,提出了一种融合三个密集连接型2D-CNNs分割结果的方法。将三维多模态MRI数据沿轴状面、冠状面和矢状面切片化,并在预处理后的切片上按比例截取33×33大小的图像块,得到三个视图的训练集;将三个训练集分别送入到密集连接型2D-CNNs模型中进行训练,得到三个分割模型;然后,将测试病人的各视图图像块依次输入到训练好的分割模型,得到脑胶质瘤三个视图的粗分割结果;将三个视图的粗分割结果进行融合处理和后处理,得到脑胶质瘤的最终分割结果,并具体划分为水肿、增强和坏死/非增强三种区域。本研究包含了BraTS2018和BraTS2013的数据集并利用Dice系数、阳性预测值、灵敏度三个指标对分割结果进行评价。实验结果表明,所提出的分割方法不仅能够精确的分割脑胶质瘤,而且可以利用多个2D-CNNs实现脑胶质瘤的三维分割。

Accurate and reliable glioma segmentation is an important prerequisite for glioma diagnosis,treatment planning,and therapeutic evaluation.In order to effectively use the characteristics of brain glioma on Magnetic Resonance Imaging(MRI)and solve the shortcomings of based on Convolutional Neural Networks(CNNs)segmentation method for brain glioma,a segmentation method is proposed that integrating three 2 D densely connected CNNs’s segmentation results.First,multimodal MRI images were sliced along the axial、coronal and sagittal views respectively.Particularly,image patches with the scale of 33*33 were cropped on the slices to obtain three training datasets.Secondly,three 2 D densely connected CNNs were trained by using three training datasets.Thirdly,the brain glioma was segmented by the three trained CNNs.Then,three segmentation results were fused.Finally,the fusion result was post-processed so that gliomas can be correctly segmented.The glioma is divided into edema tumor、necrosis/non-enhance tumor and enhance tumor;The proposed method was evaluated on the datasets of BraTS2018 and BraTS2013.At the same time,Dice score,Positive Predictive Value(PPV)and Sensitivity were used to evaluate the quality of segmentation.The experimental results showed that the proposed technique could indeed improve current segmentation accuracy.It can not only accurately complete the segmentation task of brain glioma,but also achieve 3 Dsegmentation brain glioma using multiple 2 DCNNs.