Robust Collagen Texture Quantification in Nonlinear Microscopy by Combining the Gradient Structure Tensor with a Mixed Noise Model

Collagen structure derangement is a hallmark of several diseases in different types of tissue. Extraction of quantitative markers of collagen organization from nonlinear microscopy images can help early disease detection, but it can be affected by the presence of image noise. This study introduced and validated a novel methodology for a robust quantification of collagen fiber direction, dispersion, and degree of anisotropy (DA) from collagen images, based on the reinforcement of the gradient structure tensor computation with a mixed noise model. The method was validated on a synthetic image dataset, generated by a vector field-based fiber generator algorithm, assessing its capability to distinguish different fiber organization levels and its robustness against increasing levels of image noise. The method showed an accurate estimation of fiber angle direction with errors < 2 deg for a signal-to-noise ratio (SNR) down to 5, and a slight underestimation of angle variability in the presence of fiber scales smaller than the tensor radius. Local and global DA were able to distinguish different degrees of fiber organization even at high noise level (SNR=5), with small accuracy errors for the local (<0.04) and global (<0.06) DA at realistic patterns. Proof-of-concept application of the method to human cardiac samples suggested its potential in distinguishing different collagen organization levels in real tissues. The method may represent a useful tool for the investigation of collagen structural remodeling in a variety of fibrosis-related pathological conditions. Future studies should address the method tuning to specific tissues, diseases, and clinical problems.