3.5 | Evaluation of the deep learning network
invariance with various noise distributions
Lastly, we evaluated the network’s invariancy with respect to different
noise types such as Gaussian, S&P, etc. As observed in Fig. 8, the
images in the 2nd row are corrupted with Gaussian
white noise (variance = 0.01), and the 4th row
displays S&P noise corrupted images of subcutaneous tumors. The
3rd and 5th rows are the
corresponding denoised U-Net outcomes. Our denoising U-Net efficiently
removed the Gaussian white noise whereas it is unable to reduce S&P
noise types. The reason for the good performance of the U-Net in
removing the Gaussian noise is mainly because the noise distribution of
the training images captured from the LED Acoustic-X system somewhat
looks similar to the Gaussian noise type which essentially holds the
i.i.d (independent and identically distributed) assumption. However, the
network is not effective at removing the S&P noise which needs further
investigation. Generally, the electronic noise comprises of a variety of
thermal and shot noise profiles that can be approximated to Gaussian
distribution, and thus our denoising algorithm will not have any
difficulty in handling such types of noise. Due to this capability, the
U-Net algorithm developed here has a potential of being an electronic
platform independent denoiser and further studies are under way to
evaluate this feature.