On the development of an unsupervised probabilistic algorithm for grayscale fluorescence image segmentation

Abstract: In the field of computational biology, fluorescence microscopy images often constitute the input source of information. The process of binarization of raw images to delineate interesting objects requires image segmentation into signal and background pixels. Several methods to perform image segmentation exist, the Otsu method being a popular unsupervised example. The Otsu method's lack of probabilistic predictions in terms of accuracy is a limiting factor when it comes to evaluation of the segmented images and their correctness. Based on the assumption that the background intensity distribution is Gaussian we present a new unsupervised probabilistic segmentation algorithm. The new algorithm uses Bayesian decision theory to classify pixels as signal and background respectively, and provides a prior estimate for the fractions of correctly classified pixels. Segmentation tests performed on artificial fluorescent images show that the new algorithm performs significantly better for high level of additive noise than the Otsu method. For a low level of additive noise, the new algorithm performs similarly to the Otsu method. Furthermore, the new algorithm provides a prior estimate for the fraction of correctly classified pixels close to the true values. We hope the new algorithm will constitute a good alternative to already established methods, offering precise probabilistic image segmentation.