基于多尺度卷积神经网络的磁共振成像脑肿瘤分割研究

摘要

Objective:In view of these tumors can appear anywhere in the brain and have almost any kind of shape and size, a new segmentation method of MRI based on intelligent convolutional neural network is developed.Methods: The convolutional neural network is applied in brain tumor segmentation, according to the features of the brain tumor, the multi-scale convolutional neural network is proposed and conducted multi-scale input and multi-scale down sampling to overcome the individual differences of brain tumor. At the same time it adapted any kind of size, shape and contrast of the difference layers.Results: Data from 30 patients showed that the proposed algorithm is effective. The average Dice is 83.11%, the average sensitivity coefficient is 89.48%. the average predictive positivity value coefficient is 78.91%.Conclusion: It can improve the segmentation accuracy obviously. The multi-scale convolution neural network can adaptively the differences of brain tumor, have more effective segmentation for more images.%目的：针对脑肿瘤形状、位置及大小等多变性，提出一种适合磁共振成像(MRI)脑肿瘤分割的卷积神经网络模型的改进方法。方法：将卷积神经网络应用到脑肿瘤分割上，并针对脑肿瘤的特点，提出多尺度卷积神经网络模型(MSCNN)，通过多尺度的输入与多尺度下的采样，克服脑肿瘤的个体差异，同时适应脑肿瘤不同图像层之间的大小位置差异，弱化肿瘤边缘与正常组织灰度相近的影响。结果：通过对30例患者的多模态磁共振图像进行分割，得到平均Dice系数为83.11%；平均灵敏度系数为89.48%；平均阳性预测值(PPV)系数为78.91%。结论：MRI脑肿瘤分割的改进方法可使分割精度得到明显提高，多尺度卷积神经网络能自适应脑肿瘤的差异性，并准确有效地分割脑肿瘤。