Prostate segmentation in MRI using a convolutional neural network architecture and training strategy based on statistical shape models
Prostate segmentation in MRI using a convolutional neural network architecture and training strategy based on statistical shape models
复制标题
DOI：
10.1007/s11548-018-1785-8
复制
发表时间：
2018-08-01
影响因子：
3
通讯作者：
Salcudean, Septimiu E.
中科院分区：
文献类型：
--
作者：
Karimi, Davood;Samei, Golnoosh;Salcudean, Septimiu E.
大多数现有的基于卷积神经网络（CNN）的医学图像分割方法都是基于最初为自然图像分割而开发的方法。因此，它们在很大程度上忽略了两个域之间的差异，例如目标体积的形状和外观的可变性程度较小，以及医学应用中的训练数据量较小。我们提出了一种基于CNN的方法，采用统计形状模型来解决这些问题，在MRI中的前列腺分割。方法我们的CNN预测前列腺中心的位置和形状模型的参数，这决定了前列腺表面关键点的位置。为了使用小数据训练这样一个用于分割3D图像的大模型（1），我们采用了一种分阶段的训练策略，首先训练网络以预测前列腺中心，随后添加用于预测形状模型和前列腺旋转的参数的模块，（二）我们提出了一种数据增强方法，根据计算的位移对训练图像及其前列腺表面关键点进行变形结果我们提出的方法在使用弹性网络和谱丢弃进行正则化的情况下，Dice得分达到了0.88。与标准的基于CNN的方法相比，我们的方法在前列腺基部和顶端上显示出明显更好的分割性能。我们的实验还表明，使用形状模型的数据增强显着提高了分割结果。结论先验知识的目标器官的形状可以提高基于CNN的分割方法的性能，特别是在图像特征不足以精确分割。统计形状模型也可以用来合成额外的训练数据，这可以简化大型CNN的训练。
Purpose Most of the existing convolutional neural network (CNN)-based medical image segmentation methods are based on methods that have originally been developed for segmentation of natural images. Therefore, they largely ignore the differences between the two domains, such as the smaller degree of variability in the shape and appearance of the target volume and the smaller amounts of training data in medical applications. We propose a CNN-based method for prostate segmentation in MRI that employs statistical shape models to address these issues.Methods Our CNN predicts the location of the prostate center and the parameters of the shape model, which determine the position of prostate surface keypoints. To train such a large model for segmentation of 3D images using small data (1) we adopt a stage-wise training strategy by first training the network to predict the prostate center and subsequently adding modules for predicting the parameters of the shape model and prostate rotation, (2) we propose a data augmentation method whereby the training images and their prostate surface keypoints are deformed according to the displacements computed based on the shape model, and (3) we employ various regularization techniques.Results Our proposed method achieves a Dice score of 0.88, which is obtained by using both elastic-net and spectral dropout for regularization. Compared with a standard CNN-based method, our method shows significantly better segmentation performance on the prostate base and apex. Our experiments also show that data augmentation using the shape model significantly improves the segmentation results.Conclusions Prior knowledge about the shape of the target organ can improve the performance of CNN-based segmentation methods, especially where image features are not sufficient for a precise segmentation. Statistical shape models can also be employed to synthesize additional training data that can ease the training of large CNNs.