利用多模态U形网络的CT图像前列腺分割

计算机断层扫描（computed tomography，CT）可应用于前列腺癌的检查诊断，但是它对软组织结构对比度不高，因此很难从中分割病变；而核磁共振成像（nuclear magnetic resonance imaging，MRI）具有较高的对比度，能为病变提供丰富的影像信息。为了提升CT图像的前列腺分割精度，本文提出一种新的基于深度学习的多模态U形网络图像分割模型MM-unet，充分运用MRI图像与CT图像间信息互补的特点。具体地，首先运用迁移学习思想分别训练MRI与CT图像的初始分割模型，然后通过设计一种新型的多模态损失函数MM-Loss，建立不同模态分割模型之间的联系，联合训练基于MRI与CT图像的MM-unet。为验证所提模型MM-unet的有效性，我们在某合作医院提供的Prostate数据集上进行了实验，实验结果表明，与U-net方法相比，MM-unet能够获得高出3个百分点Dice的CT图像分割精度。

Prostate segmentation in CT images with multimodal U-net

Computed tomography (CT) can be applied to prostate cancer diagnosis; however, it is not effective for the visualization of soft tissues structures because of the resulting low contrast, and thus, it is difficult to perform accurate prostate segmentation in CT images. Contrarily, nuclear magnetic resonance imaging (MRI) provides a relatively high contrast to soft tissues, which can provide rich image information for prostate segmentation. To improve the accuracy of prostate segmentation in CT images, a novel multimodal U-net (MM-unet) is proposed based on deep learning, which fully utilizes the complementary information between MRI and CT images. A transfer learning method is first applied to train the initial segmentation model parameters for segmenting MRI and CT images, and then a novel multimodal loss function MM-Loss is proposed to connect the segmentation models between different modalities, jointly training the proposed MM-unet in paired MRI and CT images. To validate the effectiveness of the proposed MM-unet, we carried out experiments on the prostate dataset provided by our allied hospital. The experimental results show that MM-unet can achieve 3% higher Dice than U-net for prostate segmentation in CT images.