There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net .
核心贡献 · Key contributions
提出具有对称收缩路径和扩展路径的 U-Net 架构，实现精确定位。 Proposed U-Net architecture with symmetric contracting and expansive paths for precise localization.
仅用 30 张标注图像，通过弹性变形数据增强在 ISBI EM 挑战中达到最先进分割性能。 Achieved state-of-the-art segmentation on ISBI EM challenge using only 30 annotated images via elastic deformation augmentation.
在 ISBI 细胞追踪挑战 2015 中以 92% IOU（PhC-U373）和 77.5% IOU（DIC-HeLa）大幅领先先前方法。 Won ISBI cell tracking challenge 2015 with 92% IOU on PhC-U373 and 77.5% on DIC-HeLa, outperforming prior methods by large margins.
引入加权损失，通过强调细胞间背景边界来分离接触物体。 Introduced weighted loss to separate touching objects by emphasizing background borders between cells.
实现 512x512 图像在 GPU 上不到一秒的快速分割。 Enabled fast segmentation of 512x512 images in under a second on a GPU.
公开提供基于 Caffe 的完整实现和训练好的网络。 Provided full Caffe-based implementation and trained networks publicly.
局限 · Limitations
需要仔细选择输入图块大小，以确保最大池化层的尺寸为偶数。 Requires careful selection of input tile size to ensure even dimensions for max-pooling layers.
无填充卷积使输出尺寸缩小固定边界宽度，限制了边缘的全上下文预测。 Unpadded convolutions reduce output size by a constant border width, limiting full-context prediction at edges.
性能依赖于弹性变形数据增强，可能无法泛化到所有生物医学任务。 Performance depends on data augmentation with elastic deformations, which may not generalize to all biomedical tasks.
训练需要足够内存的 GPU（如 6 GB）以处理大输入图块。 Training requires a GPU with sufficient memory (e.g., 6 GB) for large input tiles.
分离边界的权重图计算依赖形态学操作，增加了预处理复杂性。 Weight map computation for separation borders relies on morphological operations, adding preprocessing complexity.