3.2 | Validating U-Net architecture with out-of-class
data (photoacoustic images of in vivo mouse tumors)
The subcutaneous mouse tumors were imaged with Acoustic X system under
LED illumination. A representative schematic of the parameters used for
low frame averaged image for a certain cross-section of the tumor has
been provided. Clearly the blood vessels in the tumor that absorb the
LED illumination generate photoacoustic response that is shown with
yellow arrow in Fig. 4(a). The rightmost image denotes the outcome from
our U-Net whereas the high frame averaged image is shown at the bottom
of Fig. 4(a). Similar subcutaneous tumors have been imaged for several
mice (n=8) out of which three representative images are shown in
different columns (Fig. 4 (b-m)). The first row (Fig. 4 (b-d)) shows the
respective ultrasound images of the tumor, second row (Fig. 4 (e-g))
denotes low number of frame averaged images of the tumors, and the last
row (Fig. 4 (k-m)) has the corresponding high-number of frames averaged
photoacoustic images. Our U-Net outcomes are given in the middle row
(Fig. 4 (h-j)). The deep network model can remove the noise
significantly enough to qualitatively be similar to the high-frame
averaged image where the PA signals from the blood vessels are retained.
Given the importance of enhancing SNR in our images, we compared the
background noise distribution for the low number of frame averaged input
images with the corresponding U-Net outcome image. In Fig. S3(c), we
showed a low number of frame averaged image and the blue ROI that is
considered to calculate the background noise level. The noise level of
the same region in the corresponding U-Net provided output image was
also calculated and plotted in Fig. S3(d). Both the high-frame averaged
images and our U-Net model outcomes are visually congruent to each other
with respect to noise levels where U-Net images had 7-fold lower noise
(7.72 ± 3.07) than the low number of frame averaged image.