Efficient biomarkers for the characterization of bone tissue.

This work describes an expert system aimed to an accurate classification of cell tissue on microscopic images coming from studies of bone tissue regeneration from stem cells. We analyze a wide number of phenotype and color issues to build effective vectors of features for the subsequent characterization of tissue into five different classes: bone, cartilage, muscle, fiber and spine. The features selection includes texture, shape and color descriptors, among which we consider color histograms, Zernike moments and circular parameters. Once a preliminary set of vectors candidates are selected, several trained and non-parametric classifiers based on neural networks, decision trees, Bayesian classifiers and association rules are analyzed, and later compared with unsupervised methods to determine those that fit more closely to our needs for distinguishing bone tissue. Because of the high resolution of our biomedical images, we effectively decompose them into smaller windows for a faster execution, with the impact of the window size being discussed in terms of speed and robustness. Our final study compares accuracy and computational time together with different stainings for revealing tissue properties: Picrosirius red, alcian blue and safranin blue. Overall, safranin blue reveals as the best staining and multilayer perceptron as the most effective classifier.