**5. Conclusions**

Automated recognition of tooth-marked tongue would not only have benefit in clinical settings, but it would also be invaluable in the delivery of health care to populations with inadequate access to diagnostic imaging specialists. In this paper, we have presented a tooth-marked tongue recognition method based on deep features and localized discriminative regions using Grad-CAM. The experiments showed that the proposed method has greatly improved accuracy, compared to previous methods. We mainly improve the problem in the following three aspects: (1) Analysis of the entire tongue picture—no need to cut small patches. (2) Analyzing the influence of different receptive field sizes on the classification results and finding the receptive field size suitable for the tooth-marked tongue problem. (3) Enhancing the interpretability of the CNN algorithm. Future work includes two aspects: (1) More new tongue samples will be acquired. Since we use a deep CNN as a feature extractor, the proposed model will benefit a lot from a larger dataset. (2) We will improve the architecture of the CNN model to further improve the accuracy and reduce the computation cost, and use patient history (or other relevant characteristics) in the future.

**Author Contributions:** Methodology, Experimental analysis and Paper Writing, Y.S.; Writing-review and Data analysis, S.D.; Data and Writing Correction, Y.Z. and J.L.; The work was done under the supervision and guidance of X.L.; All the authors revised the manuscript.

**Funding:** This research received no external funding.

**Conflicts of Interest:** The authors declare no conflicts of interest.

## **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
