Fig. 6.

Evaluation of noise invariancy for top-performing deep learning networks: (a) and (e): showcase images corrupted with Gaussian white noise (variance = 0.01) and S&P noise, derived from a subcutaneous mouse tumor cross-section. The outcomes produced by (b) and (f): UN, (c) and (g): Dense-UN, and (d) and (h): R2-UN architectures demonstrate their robustness to various noise distributions. Last row: PSNR comparison for the networks’ reconstructed images whose input was corrupted by (i) Gaussian and (j) S&P noise. SSIM for networks’ outcomes whose input was adulterated by (k) Gaussian and (l) S&P noise. (#: p < 0.001, ns: Not Significant).