Performance evaluation of a robust fuzzy noise estimation technique in MRI

Noise estimation is a critical component in MRI image retrieval systems. Despite extensive studies, there are several major shortcomings with the existing techniques for noise estimation learning that can significantly affect their application to MRI image diagnosis. This paper focuses on a novel pixel wise fuzzy based noise evaluation that goes beyond the conventional noise estimation approaches with respect to visual resemblance and original image annotations. Based on this framework, a Fuzzy Image Noise Ratio (FINR) has been concocted for MR Images corrupted by additive White Gaussian Noise. An improvement in FINR can be translated into an increase in image acquisition speed and therefore be used to reduce the imaging cost and motion artifacts. Also, since an improvement in FINR can significantly reduce imaging times, it can increase the cost-effectiveness of MRI equipment in a hospital environment. The robustness of the proposed FINR in quantization of noise has been validated by evaluating its performance on 5 MR Brain image sets contaminated by Gaussian noise of 30% and 70% variances. The results have been compared with the performance of existing state-of-the-art metrics. In addition a partition based FINR evaluation on different segments of MRI image has been computed to show the noise correlations between different similar and dissimilar parts.