Table S1 - An Automated Blur Detection Method for Histological Whole Slide Imaging

Characteristics of the DT models produced during the training phase. For each training set of tiles we ranked the DT models according to their respective accuracy. Column "Set" gives the name of the dataset used (HE  =  hematoxylin and eosin, NS  =  nuclear staining, CS  =  cytoplasmic staining, IHC  =  immunohistochemistry, ALL  =  complete training set) and column "N" the number of tiles in the dataset. Column “DT-Algorithm” indicates the split criterion used to create the DT (Discri multi  =  discriminant-based multivariate split, Discri uni  =  discriminant-based univariate split, Gini  =  exhaustive search for univariate split evaluated by the Gini index of node impurity). Column “Features” lists the set of features used for training (HC  =  Haralick contrast, MGM  =  mean gradient magnitude, TG  =  Tenengrad function, NO  =  noise, HE  =  Haralick entropy, SDBD  =  standard deviation of blur difference, SH  =  sharpness, MBD  =  mean blur difference). Columns "B -> S" and "S -> B" show the number of false negative decisions (i.e., the blurred tiles classified as being sharp) and the number of false positive decisions (i.e., the sharp tiles classified as being blurred), respectively. The classification performances provided in column “Accuracy” resulted from a nested cross-validation (5-fold x 5-fold) on each training set. For each dataset, we examined the DT models whose accuracy was higher than the maximum accuracy minus 0.005. DTs satisfying that condition are surrounded by a bounding box. For these models, additional columns detail the number of nodes in the tree (N nodes in tree), the number of tests in the tree (N tests in tree), the maximum depth of the tree (Tree depth), the number of features selected in the final DT model and the corresponding set of ignored features. The DT model finally selected is outlined in green (see main text for details).
(PDF)