Table 2 Comparison of eight classifiers for 2D and 3D image classification. The average classification accuracy on test data (from 10-fold cross-validation) is shown for the optimal parameters settings (shown in parentheses) for each classification approach. The parameters are: nhu – number of hidden nodes in neural network, stop-fract – the fraction of the training data used to stop neural network training, C – error penalty in SVMs, sigma – kernel variance in SVMs, nboost – total number of iterations in AdaBoost, nbag – total number of iterations in Bagging, nhug – number of hidden nodes in the gating network of Mixtures-of-Experts, and ne – total number of experts in Mixtures-of-Experts. The accuracies across the 10-fold cross-validation trials were compared to those for the previously described neural network configuration (nhu = 20, stop-fract = 0.3) by a paired t-test (88% for SLF13, 86% for SLF8, 93% for SLF10, and 84% for SLF14). The best performances are underscored and highlighted for each feature set. *CPU times listed for each classifier are for training and testing for all images in each cross-validation trial (training times include times calculating features), which were measured on an Athlon 1.7 GHz processor with 1.5 GB memory running Redhat Linux 7.1.

From: Boosting accuracy of automated classification of fluorescence microscope images for location proteomics