Fig 1.
Histological image samples: From left to right, normal tissue followed by cancer of grades G1, G2 and G3.
Fig 2.
Organization of the experiments.
Table 1.
Results on the IMEDIATREAT data set.
Classification accuracy and confusion matrix in percents (true on the columns and predicted on the rows) for the combined approach.
Table 2.
Statistical measurements per class in percents for the IMEDIATREAT data set for the combined approach.
Table 3.
Classification test accuracies in percents for the individual classifiers and for each data set in turn.
Fig 3.
Box plots showing the variation in weight for each classifier in turn as determined by the final population of the DE.
First row corresponds to the weights found for IMEDIATREAT and GlaS (test set B) data sets, second row for 40x, 100x for BreaKHis, while the last row show the values for 200x and 400x respectively. The blue filled circle represents the best configuration result.
Fig 4.
Correlations between outputs of the individual classifiers.
On the left, Pearson product-moment correlation coefficients takes into account probabilities outputs, while on the right Cohen’s kappa coefficients are based on the actual test accuracy results. Darker colors correspond to higher correlation.
Table 4.
Comparison on the benchmark data sets.
Applications to GlaS and BreaKHis data sets and comparison to results obtained by other techniques.
Fig 5.
Test classification accuracy for BreaKHis when using weights discovered by DE as applied to the IMEDIATREAT data set, and the gains in percent when the DE optimizes the weights directly on BreaKHis.
Left plot contains image level recognition, while the right one illustrates patient level recognition.
Fig 6.
Sample images that are wrongly classified by the proposed classifier.
Their assigned grades are G0 (image from the first row and first column), G1 (first row, second column) and G2 for the rest (since G2 images are those that are most commonly misclassified).
Fig 7.
The confidence threshold is varied from 0.5 to 1.
The trusted failures from the first row right plot are included in the trusted results from the first row left plot. Analogously, unreliable failures are part of the unreliable results. The plot from the second row illustrates the Pareto front between the trusted samples (denoted by “T” in the labels) and the trusted but wrongly classified (denoted by “TW”): both axes are in terms of percents.