**5. Discussions**

Our experimental results empirically demonstrated that DL reconstruction in certain situations is superior to the conventional iterative reconstructions, especially when it deals with signals compromised by strong noise. The test on the sound reflectors suggests a grea<sup>t</sup> potential of DL-based methods for removing reflection artifacts in the PAT image formation.

However, there are some limitations in this study. First, some of the parameters in these iterative algorithms, such as the number of iterations and the value of the regularization parameter, are not specifically optimized. Varying these parameters may result in somewhat improved performance of the iterative algorithms. However, finding the optimal values of such parameters are highly empirical, and there is no universal approach in general. Second, we empirically used the combination of MSE and SSIM as the loss function to optimize the network, and evaluated the image quality by PSNR and SSIM. The choice of the loss function and image metrics may not be the optimal to capture the visual quality for PAT imaging. Third, we only studied the reflection artifacts under different noise levels and a simple variable sound model, other more complicated conditions such as limited views and more complicated variable sounds can be surely addressed by extending the proposed network.

*Appl. Sci.* **2019**, *9*, 2615

**Figure 6.** The box plots of the image quality evaluated by PSNR and SSIM on the testing set in a heterogeneous medium. From top row to bottom row: noise increases from 0 to 0.4. Diamonds () indicate outliers.

**Figure 7.** Reconstruction in a heterogeneous Medium.
