Drag the slider to vary the level of noise added to the images and observe how the two methods respond differently to increasing noise. You can also zoom, pan, and vary the noise level with the mouse wheel.
This is the standard approach for parameter estimation.
Without added noise, the results in the prostate look good, but you can see some noise amplification in low SNR regions like the muscles and the rectum. As the noise level increases the maps amplify the noise further and are eventually uninterpretable.
Without added noise, the CNN gives similar results as NLLS fitting in the prostate, which has high SNR. At lower levels of added noise, the CNN continues to provide good maps in high-SNR regions at the cost of some blurring. At higher noise levels the CNN gives maps with unusable levels of blurring, warping, and hallucinations.
Consistent with the quantitative analyses in the paper, the CNN and NLLS methods give similar results at high SNR, and the CNN outperforms the NLLS fit in lower SNR levels. At very low SNR levels ($sigma$ > ~0.06) both methods fail.