前列腺磁共振图像分割的反卷积神经网络方法

发布时间：2018-04-13 05:04

本文选题：前列腺分割 + 磁共振图像　；参考：《中国图象图形学报》2017年04期

【摘要】：目的前列腺磁共振图像存在组织边界对比度低、有效区域少等问题,手工勾勒组织轮廓边界的传统分割方法无法满足临床实时性要求,针对这些问题提出了一种基于深度反卷积神经网络的前列腺磁共振图像分割算法。方法基于深度学习理论,将训练图像样本输入设计好的卷积神经网络,提取具有高度区分性的前列腺图像特征,反卷积策略用于拓展特征图尺寸,使网络的输入尺寸与输出预测图大小匹配。网络生成的概率预测图通过训练一个softmax分类器,对预测图像取二值化,获得最终的分割结果。为克服原始图像中有效组织较少的问题,采用dice相似性系数作为卷积网络的损失函数。结果本文算法以Dice相似性系数和Hausdorff距离作为评价指标,在MICCAI 2012数据集中,Dice相似性系数大于89.75%,Hausdorff距离小于1.3 mm,达到了传统方法的分割精度,并且将处理时间缩短在1 min以内,明显优于其他方法。结论定量与定性的实验表明,基于反卷积神经网络的前列腺分割方法可以准确地对磁共振图像进行分割,相比于其他分割算法大幅度减小了处理时间,能够很好地适用于临床的前列腺图像分割任务。
[Abstract]:Objective there are some problems in prostatic magnetic resonance imaging such as low contrast of tissue boundary and less effective area. The traditional segmentation method based on manual delineation of tissue contour can not meet the requirement of clinical real-time.To solve these problems, a prostatic magnetic resonance image segmentation algorithm based on deep deconvolution neural network is proposed.Methods based on the theory of depth learning, the training image samples were input into the designed convolution neural network, and the prostate image features were extracted, and the deconvolution strategy was used to expand the size of the feature image.The input size of the network is matched with the size of the output prediction chart.By training a softmax classifier, the probabilistic prediction graph generated by the network is binarized to the predicted image and the final segmentation result is obtained.In order to overcome the problem of less efficient organization in the original image, the dice similarity coefficient is used as the loss function of the convolution network.Results the Dice similarity coefficient and the Hausdorff distance are taken as the evaluation indexes in this algorithm. In the MICCAI 2012 data set, the similarity coefficient is greater than 89.75 and the Hausdorff distance is less than 1.3 mm, which achieves the segmentation accuracy of the traditional method, and reduces the processing time to less than 1 min.It is obviously superior to other methods.Conclusion quantitative and qualitative experiments show that the prostate segmentation method based on deconvolution neural network can accurately segment the magnetic resonance image, which greatly reduces the processing time compared with other segmentation algorithms.It is suitable for clinical prostate image segmentation.
【作者单位】：合肥工业大学计算机与信息学院;安徽省医科大学第二附属医院泌尿外科;
【基金】：国家自然科学基金项目(61371156)~~
【分类号】：TP391.41;TP183

【相似文献】