New resolution independent approach to noise estimation in Minimum Noise Fraction denoising of tissues measured with Infrared Imaging

Abstract

The Minimum Noise Fraction (MNF) was originally developed for remote sensing data denoising. However, due to the hyperspectral character of Infrared (IR) imaging data, MNF is becoming increasingly popular in this field. IR imaging obtains data rich in biochemical information, therefore, it is extensively applied in tissue analysis and classification. It is popular to use different IR objectives of different magnification, which combined with Focal Plane Array (FPA) detector of fixed pixels size, give varying projected pixel sizes and as a result – different spatial resolutions. The MNF method proposed by Green et al. uses a noise correlation matrix that is calculated based on neighboring pixel difference, which heavily depends on the registered image magnification and thus spatial resolution, leading to different denoising efficiencies and signal loss. Here, we presented a new resolution-independent approach to noise matrix estimation (MNF2) based on the 2 sequential scans of the sample. In this way, the data acquisition step is only minimally changed, while offering a uniform denoising step. We decided to test our method on representative breast tissue biopsy, due to its spatial and biochemical heterogeneity. We compared the performance of the new method for data measured using objectives giving different spatial pixel projections: low - 11.4 μm (3.5x), standard - 2.7 μm (15x), and high - 1.1 μm (36x). Results of this study show that conventional MNF for standard and low resolutions produces an altered denoised signal, while MNF2 gives a denoised signal of very good quality. In the case of high resolution data, standard MNF gives reasonable results, likewise proposed MNF2.