FIGURE 3 (a) Photograph of the 3D phantom and the 2D
cross-section view in yellow dotted line whose scanning plane is imaged
with LED-based PA system. (b) The corresponding US image. (c, d) Low
number of frame averaging images in 2D and 3D view. (e, f) Corresponding
high number of frame averaging images in 2D and 3D view. (g, h) The
U-Net outcomes in 2D and 3D view.
We calculated various quantitative metrics to measure the enhancement of
image quality in the U-Net images over the low number of frame averaged
image inputs. In Fig. 5 (a) and 5 (e), a representative US image and a
low-averaged photoacoustic image of in vitro phantom (graphite rod
embedded into gelatin base) and in vivo subcutaneous tumor are
shown where the orange ROI box denotes the background, and the green ROI
box signifies the target signal ROI.
We considered three image quality metrices namely SNR, PSNR and CNR for
the analysis. There was a statistically significant difference for all
the metrics between the U-Net and low frame average images (Student’s
t-test with 99% confidence level (α = 0.01)). The SNR improvement
achieved with U-Net architecture as compared to the low number of frame
averaging is approximately 4.39 ± 2.55-fold. Similar to Fig. 5 (a-d),
the data calculated from 24 mouse tumor images (8 mice x 3
cross-sections/mouse) is shown in Fig. 5 (e-h). All the metrices were
calculated considering various tumor ROIs and background zones. There is
approximately a 5-fold increase (4.27 ± 0.87) in SNR between low number
of frames averaging and the U-Net outcomes. Clearly, we also notice no
significant difference in the image quality metrics observed between the
U-Net and the high number of frame averaged images.