Elamur Research
☰
home / papers / 2015 / U-Net: Convolutional Networks for Biomedical Image Segmentation

U-Net: Convolutional Networks for Biomedical Image Segmentation

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "arXiv — non-exclusive license to distribute" does not allow redistributing the text itself. Original — arxiv.org/pdf/1505.04597

Summary

Biomedical image analysis often requires a class label for every pixel, yet annotated training images are scarce. Earlier sliding-window CNNs were slow and had to trade localization precision against context. The authors build on fully convolutional networks and propose U-Net, a symmetric encoder–decoder. A downsampling branch gathers context. An upsampling branch with many feature channels restores resolution. At each level, cropped high-resolution features from the downsampling side are concatenated into the upsampling side. The model uses only valid convolutions and has no fully connected layers, so arbitrarily large images can be processed tile by tile, with mirrored borders. To cope with tiny datasets, training relies heavily on random elastic deformations. A per-pixel weighted loss emphasizes the thin background gaps between touching cells. Trained end-to-end in Caffe from only a few dozen images, U-Net set a new best warping error on the ISBI EM segmentation challenge. It also clearly won two ISBI 2015 cell tracking datasets. The work showed that accurate dense segmentation is feasible with very little labeled data, and the architecture became a widely reused template.

Key points

  • Architecture: 23 convolutional layers built from unpadded 3x3 convolutions with ReLU, 2x2 max pooling, and 2x2 up-convolutions, plus a final 1x1 classifier. Weights use Gaussian (He-style) initialization. Training uses SGD with batch size 1 and momentum 0.99.
  • Augmentation and loss: Elastic deformations are generated from random displacements on a coarse 3x3 grid (std 10 px) with bicubic interpolation. The loss weight map uses w0 = 10 and σ ≈ 5 px.
  • ISBI EM challenge (30 training images, 512x512, Drosophila VNC): warping error 0.000353 and Rand error 0.0382. The previous IDSIA sliding-window CNN scored 0.000420 and 0.0504.
  • ISBI cell tracking challenge 2015 (IOU): PhC-U373 (35 images) reached 92% versus 83% for the runner-up. DIC-HeLa (20 images) reached 77.5% versus 46%.
  • Efficiency: A 512x512 image is segmented in under a second on a GPU. Training takes about 10 hours on an NVidia Titan (6 GB). The Caffe code and trained models were released.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Other

Cited by

714 papers of our library cite it (of the 140673 whose bibliographies we read); 6 of them have a page here, 708 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1505.04597 · arXiv 1505.04597

@misc{ronneberger2015u,
  title = {{U-Net: Convolutional Networks for Biomedical Image Segmentation}},
  author = {Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas},
  year = {2015},
  doi = {10.48550/arxiv.1505.04597},
  eprint = {1505.04597},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1505.04597}
}

References

14 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ciresan, D.C., Gambardella, L.M., Giusti, A., Schmidhuber, J.: Deep neural networks segment neuronal membranes in electron microscopy images. In: NIPS. pp. 2852–2860 (2012) find
  2. Dosovitskiy, A., Springenberg, J.T., Riedmiller, M., Brox, T.: Discriminative unsupervised feature learning with convolutional neural networks. In: NIPS (2014) find
  3. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2014) find our card
  4. Hariharan, B., Arbelez, P., Girshick, R., Malik, J.: Hypercolumns for object segmentation and fine-grained localization (2014), arXiv:1411.5752 [cs.CV] — [arXiv:1411.5752](https://arxiv.org/abs/1411.5752) find
  5. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification (2015), arXiv:1502.01852 [cs.CV] find
  6. Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadarrama, S., Darrell, T.: Caffe: Convolutional architecture for fast feature embedding (2014), arXiv:1408.5093 [cs.CV] find our card
  7. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: NIPS. pp. 1106–1114 (2012) find
  8. LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E., Hubbard, W., Jackel, L.D.: Backpropagation applied to handwritten zip code recognition. Neural Computation 1(4), 541–551 (1989) find
  9. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation (2014), arXiv:1411.4038 [cs.CV] — [arXiv:1411.4038](https://arxiv.org/abs/1411.4038) find
  10. Maska, M., (...), de Solorzano, C.O.: A benchmark for comparison of cell tracking algorithms. Bioinformatics 30, 1609–1617 (2014) find
  11. Seyedhosseini, M., Sajjadi, M., Tasdizen, T.: Image segmentation with cascaded hierarchical models and logistic disjunctive normal networks. In: Computer Vision (ICCV), 2013 IEEE International Conference on. pp. 2168–2175 (2013) find
  12. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2014), arXiv:1409.1556 [cs.CV] find our card
  13. WWW: Web page of the cell tracking challenge, http://www.codesolorzano.com/celltrackingchallenge/Cell\_Tracking\_Challenge/Welcome.html ↗
  14. WWW: Web page of the em segmentation challenge, http://brainiac2.mit.edu/isbi\_challenge/U-net implementation, trained networks and supplementary material available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net ↗