Transferable Visual Words: Exploiting the Semantics of Anatomical Patterns for Self-supervised Learning

We have introduced a new concept called transferable visual words (TransVW), where the sophisticated, consistent, and recurring anatomical patterns across medical images are natural visual words. These visual words, associated with rich semantics about the human anatomy, can be automatically extracted from unlabeled medical images and serve as strong yet free supervision signals for CNNs to learn generic semantics-enriched representation via self-supervision. TransVW is a fully autodidactic scheme (requiring no manual labeling), serves as an add-on strategy (boosting existing self-supervised methods), and yields semantics-enriched pre-trained models (saving annotation efforts for a variety of 3D medical applications through transfer learning).

Paper

Transferable Visual Words: Exploiting the Semantics of Anatomical Patterns for Self-supervised Learning

Fatemeh Haghighi¹, Mohammad Reza Hosseinzadeh Taher¹,Zongwei Zhou¹,Michael B. Gotway², Jianming Liang^1
1 Arizona State University, ² Mayo Clinic
IEEE Transactions on Medical Imaging (TMI)
paper | code

Available implementation

                             

---

Visual word learning is an add-on strategy, which boost existing self-supervised methods.

TransVW offers superior transfer learning performance

TransVW accelerates the training process

TransVW reduces the annotation cost

Credit to superbar by Scott Lowe for Matlab code of superbar.

Citation

If you use our source code and/or refer to the baseline results published in the paper, please cite our paper by using the following BibTex entry:

@ARTICLE{haghighi2021transferable,
  author={Haghighi, Fatemeh and Taher, Mohammad Reza Hosseinzadeh and Zhou, Zongwei and Gotway, Michael B. and Liang, Jianming},
  journal={IEEE Transactions on Medical Imaging}, 
  title={Transferable Visual Words: Exploiting the Semantics of Anatomical Patterns for Self-Supervised Learning}, 
  year={2021},
  volume={40},
  number={10},
  pages={2857-2868},
  doi={10.1109/TMI.2021.3060634}}

Acknowledgement

This research has been supported partially by ASU and Mayo Clinic through a Seed Grant and an Innovation Grant, and partially by the NIH under Award Number R01HL128785. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. This work has utilized the GPUs provided partially by the ASU Research Computing and partially by the Extreme Science and Engineering Discovery Environment (XSEDE) funded by the National Science Foundation (NSF) under grant number ACI-1548562. We thank Zuwei Guo for implementing Rubik's cube, M. M. Rahman Siddiquee for examining NiftyNet, and Jiaxuan Pang for evaluating I3D, Shivam Bajpai for helping in adopting TransVW to nnU-Net, and Shrikar Tatapudi for helping improve the writing of this paper. The content of this paper is covered by patents pending.

License

Released under the ASU GitHub Project License.