**4. Conclusions**

In this research, we explored the feasibility of characterizing adipocyte size in human adipose tissue using PA measurement combined with a deep learning method. In the experiments on *ex vivo* human adipose tissues, we first studied different networks with various numbers of layers and the results indicated that the network depth has a grea<sup>t</sup> influence on performance of network. As shown in Figure 4, the 7-layer network yielded the best performance and resulted in a nearly 10% decrease of MRE compared with the 2-layer network when training on the 12.5 to 16.5 MHz spectral band. When trained on different spectral bands and the entire spectral band, our network found out the optimally fitted relationship between the PA signal spectrum and the adipocyte size measured by histology. The most sensitive spectral band turned out to be relative to adipocyte size which can be evaluated by Equation (5). PA signals are characterized with a broad-band property; combining all spectral components resulted in an MRE decrease of 3.48% compared with training on the most sensitive spectral band in our study. Compared to conventional PASA method, using a deep learning method to fit the nonlinear relationship can better utilize the rich information in the power spectrum and an improvement of 12.84 % on MRE was achieved. The results of our research show the validity of analyzing PA signals in the frequency domain using deep learning which can be a novel method for quantitative and noninvasive evaluation of biological tissues, e.g., characterizing human adipose tissue cellular phenotype in the context of the clinical managemen<sup>t</sup> of obese patients. In our further study, a larger data set will be built and a method of combining multiple wavelengths, instead of just using 1210 nm laser illumination, will be tested to improve accuracy of evaluation for clinical application.

**Author Contributions:** Conceptualization and methodology, J.Y. and X.W.; Data Curation, M.C., T.F., A.R.W., N.A.B., C.N.L. and R.W.O.; Software, X.M. and C.M.; Formal Analysis, X.M.; Writing, X.M. and C.M.; Supervision, Q.S., J.Y., Q.C. and X.W.

**Funding:** This research was funded by the National Key Research and Development Program of China (No. 2017YFC0111402), the Natural Science Funds of Jiangsu Province of China (No. BK20181256), and by NIH grants R01DK115190 (RWO, CNL) and R01DK097449 (RWO).

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