Precise segmentation of the visual pathway is significant in preoperative planning to prevent the surgeon from touching it during the operation. Manual segmentation is time consuming and tedious. Thus, automatic segmentation strategies are necessary to assist clinical evaluation. However, the low contrast and blurred boundary between the target and the background in the image make automatic segmentation a challenging problem. This paper proposed a spatial probabilistic distribution map (SPDM)-based two-channel 3D U-Net to make shape and position prior information available for deep learning. First, an atlas calculated by group-wise registration was used to register each training volume image for deformation field determination. Second, the deformation field was used to transform the label of the corresponding training image to the template space, and then all the warped labels were summed up to create an SPDM. Third, the region of interest of the image and SPDM were sent to the network to predict the final segmentation. The proposed method was evaluated and compared against a conventional 3D U-Net on two datasets. Experimental results indicated that our method overcame the problem of low contrast and achieved better performance than previous methods.

视觉通路的精确分割在术前计划中很重要，可防止手术过程中外科医生触摸它。手动分割既费时又乏味。因此，自动分割策略对于协助临床评估是必要的。然而，图像中目标和背景之间的低对比度和模糊边界使自动分割成为一个难题。本文提出了一种基于空间概率分布图（SPDM）的两通道3D U-Net，以使形状和位置先验信息可用于深度学习。首先，使用通过逐组配准计算出的地图集配准每个训练体积图像，以确定变形场。其次，使用变形场将相应训练图像的标签转换为模板空间，然后汇总所有变形的标签以创建SPDM。第三，将图像的感兴趣区域和SPDM发送到网络以预测最终的分割。对所提出的方法进行了评估，并与两个数据集上的常规3D U-Net进行了比较。实验结果表明，我们的方法克服了对比度低的问题，并且比以前的方法具有更好的性能。