Blind Color Deconvolution and Classification of Histological Images Using the Hyperbolic Secant Prior

In this paper, we present a novel approach to Blind Color De-convolution, a stain separation technique useful for normalizing, augmenting, and automatically diagnosing histological images. To robustly estimate the stain colors and concentrations, we follow the Bayesian framework and introduce a Gaussian prior distribution on the color vectors and a Hyperbolic Secant prior on the concentrations, which is a seldom explored image model.We provide a comprehensive mathematical derivation of our inference procedure, outlining the underlying principles and assumptions. To demonstrate its effectiveness, we conduct two experiments. The first experiment assesses the fidelity of the reconstructed images to the underlying tissue characteristics. The second experiment uses it as a preprocessing step for a multicenter breast cancer classification task. Our results reveal the superior performance of the proposed method, underscoring its potential for advancing histological image processing and AI-assisted diagnosis.