3.3 | Comparison of U-Net architecture outcome with
traditional noise removal algorithms
Our U-Net architecture is compared to the outcomes generated by the
traditional algorithms like Wiener, Median, TV and SG filtering that
were not reported in previous studies involving deep learning
technologies and LED based photoacoustic images [54, 55]. The result
of the subcutaneous tumor images of three different mice applying all
those algorithms along with the low number of frame averaged inputs is
shown in Fig. 6. It is to be noted that the varying amount of noise for
each of the input images was intentionally chosen to evaluate noise
invariancy of our U-Net algorithm with respect to the noise levels. The
noise level for mouse 2 is the lowest among the three data sets shown
and that of mouse 3 is highest (achieved by varying the gain range). The
images for mice #1, #2, and #3 were acquired with 22-, 19-, and 25-dB
gain, respectively. We did a quantitative comparison of all the
different noise removal algorithms along with our U-Net and the box
plots for PSNR, SNR and CNR are shown in Fig. 6 (h), 6 (p), & 6 (x),
respectively. When we compared the performance of our model network with
the traditional noise removal algorithms like Wiener, Median, TV, and SG
filter, the U-Net images produced a better-enhanced version of outputs
than the other ones which can visually be confirmed from Fig. 6 (b-g,
j-o, and r-w). The different noise levels in Fig. 6 (b), (j), & (r) did
not affect our network outcomes, which makes the U-Net framework a
noise-level invariant model, while other established algorithms are not
robust enough to deal with different kinds of noise profiles. In
comparison with the other algorithms, U-Net emerges not only as a
noise-robust process, but it maintains the original phantom structure
that other methods are incapable of achieving. Wiener and SG filters
blur the images along with SG showing an artifact similar to hazy motion
artifact during acquisition. There is not much improvement in the signal
and contrast concerning the background noise for the Median filter, a
traditionally used noise filter. The TV-L1 filter reduced the noise
homogeneously, however only the brightest parts of the images remained
intact and a significant enhancement in background occurred in the
entire image.