UNetPlusPlus: two papers, two source directories, and a citation that stops mid-author

The second BibTeX entry ends at the third author

The citation section is the only place this repository prints anything close to code, and the block is unfinished. The first entry, keyed zhou2019unetplusplus for the IEEE Transactions on Medical Imaging paper, carries every field a citation needs: journal, volume 39, number 6, pages 1856 to 1867, year 2020, doi 10.1109/TMI.2019.2959609, publisher IEEE, and all four authors. The second entry, keyed zhou2018unetplusplus as an incollection for the 2018 Deep Learning in Medical Image Analysis paper, gives a title and then stops three names into the author list, at Tajbakhsh, Nima. No closing brace follows it. There is no venue, no year, no pages, no DOI, no publisher. Paste that block into a bibliography file as printed and a reference manager will trip over the entry that never terminates. The dropped fields are exactly the ones a citation to the earlier paper needs, so rebuilding it means going back to the source document rather than to this file. Worth noting on the complete entry: the DOI carries a 2019 prefix while the year field says 2020, the ordinary gap between acceptance and the issue landing.

Two papers share one code link, and nothing maps the directories to either

The architecture is presented twice under two different titles from the same author group at Arizona State University. The 2018 conference version, A Nested U-Net Architecture for Medical Image Segmentation, is marked Oral and carries links to slides, a poster, and a blog post in addition to the paper and code. The journal version, Redesigning Skip Connections to Exploit Multiscale Features in Image Segmentation, carries only a paper link and a code link. Both code links point at this one repository, and neither paragraph states which of the two papers the keras/ and pytorch/ trees implement. The titles also signal different emphases, one on the nested arrangement of U-Nets at varying depths and one on the dense decoder wiring at a shared resolution. The repository description adds a third piece of metadata, a claim of an IEEE TMI Best Paper Award, which lines up with the journal entry and its 2020 issue placement rather than with the 2018 conference paper.

The official implementation is a two-line directory list

The Official implementation section contains two lines, keras/ and pytorch/, and that is the whole entry path offered. There is no pip command, no requirements file, no framework version, no class name, no function signature, no pretrained weight download, and no usage snippet anywhere in the file. The top-level entries match that absence: .gitignore, CITATION.cff, Figures/, LICENSE, README.md, keras/, pytorch/, with no setup.py, no pyproject.toml, and no requirements.txt at the root, even though Python is the primary language. What lands in your checkout is source to read and adapt, not a package to install and pin.

No releases leaves the commit as the only version marker

The repository has no GitHub releases, so there is no tagged artifact, no changelog, and nothing to diff a later snapshot against. The last recorded push is 2026-08-25 on the master branch, recent enough that the code is plainly not abandoned, but recent commits do not stand in for a version number. Combined with the absent dependency manifest at the root, the framework you build against is whatever already sits in your environment. If you need to reproduce a published result from this code, record the commit hash yourself at the start, because the project will not do it for you.

License metadata, a LICENSE file, and a patent-pending line

The license metadata recorded for this repository reads NOASSERTION, while a LICENSE file sits in the root. The two do not agree, and nothing here says which text governs. The Acknowledgments section adds a constraint that matters more for a medical imaging method than the file name does: the work was supported in part by NIH Award Number R01HL128785 and by Arizona State University and Mayo Clinic through a Seed Grant and an Innovation Grant, and the closing sentence states that this is a patent-pending technology. Read the LICENSE file and the award terms directly before any of this code reaches a product.

The answer to needing a runnable example points at other maintainers

Five other implementations are listed, four in PyTorch and one a Keras kernel on Kaggle that pairs the nested U-Net with an EfficientNet encoder. One entry is marked deprecated in place, the MontaEllis Pytorch-Medical-Segmentation repository, which leaves qubvel-org/segmentation_models.pytorch, 4uiiurz1/pytorch-nested-unet, ZJUGiveLab/UNet-Version, and Siddhartha's kernel as the live references. The tradeoff is stated plainly by the list itself: the official repository ships no example, and the first working code a reader finds is maintained by somebody else.

One paragraph of architecture prose carries the whole description

Setting the citation block aside, the descriptive text amounts to a single paragraph. UNet++ is a nested U-Net architecture for medical image segmentation, built from U-Nets of varying depths whose decoders are densely connected at the same resolution through redesigned skip pathways, aimed at two named limitations of the original U-Net, the unknown optimal depth of the architecture and the unnecessarily restrictive design of its skip connections. Everything else is links. Even the header is HTML rather than Markdown, a centered h1, a badge whose alt text is empty, and an empty paragraph element, with the illustrative material held in a root Figures/ directory.

Editorial conclusion

Treat this repository as paper reference code, not a dependency. Anyone planning to build on UNet++ should read the two framework directories, record a commit hash since there are no releases, settle the license yourself because the metadata says NOASSERTION while a LICENSE file exists, and check the patent-pending note before shipping anything clinical.

Frequently asked questions

Official sources

1. Issues
2. MrGiovanni/UNetPlusPlus on GitHub
3. Project website
4. README