**6. Conclusions**

In this article, we have proposed a novel DNN to remove the reflection artifacts in reconstructed PAT images under different noise levels and different media. By directly comparing the proposed network to popular iterative reconstruction algorithms with simulated PAT data from CT scans, the results have showed that the proposed network is able to reconstruct superior images over the conventional iterative reconstructions in typical scenarios in terms of computational efficiency and noise reduction.

The results can be further strengthened in several aspects. One practical and significant question is how to make the network robust to potential malignant attacks. It is well known that DL models are vulnerable to adversarial examples. A more stable training procedure is thus critical to the clinical application of DL methods. Next, the effectiveness and efficiency of the network can be further improved for constrained resources and cloud-end processing. Some other factors, such as limited view, acoustic attenuation, and fluctuation of sound speed, can greatly impact the quality of PAT images. It would be interesting to extend the DL approach to these situations as well.

**Author Contributions:** H.S. and Y.Y. initiated the project and designed the experiments. H.S. performed machine learning research. Y.Y. performed iterative reconstruction research. H.S. and Y.Y. wrote the paper, and G.W. participated in the discussions and edited the paper.

**Funding:** Y.Y. was partly supported by the NSF gran<sup>t</sup> DMS-1715178, the Simons travel grant, and the startup fund from the Michigan State University.

**Acknowledgments:** The authors thank NVIDIA Corporation for the donation of GPUs used for this research. The authors would also like to express their gratitude to the anonymous reviewers for the valuable suggestions and comments which helped considerably improve the exposition of the paper.

**Conflicts of Interest:** H.S. and G.W. have received unrelated industrial research grants from General Electric and Hologic Inc.
