Fig 1.
Examples of misclassified pixels due to majority voting-based labelling in SICAPv2 dataset.
The pixels belonging to the minority class don’t get acknowledged separately (a) RGB Patch; (b) Labelling Mask; (c) Probability distribution estimate of the grades/classes represented in the Mask.
Fig 2.
Probability distributions of misclassified pixel belonging to different classes in SICAPv2 dataset partition ‘Val1’ (a) G3 training; (b) G3 test; (c) G4 training; (d) G4 test; (e) G5 training; (f) G5 test.
Table 1.
Summarized statistics related to misclassification of different classes in SICAPv2 dataset partitions.
Fig 3.
Flowchart of the proposed ensemble classifier with individual CNN-based one-vs-all binary classifiers and multi-label output.
Fig 4.
Training and testing paradigm for the proposed multi-label ensemble classifiers.
Fig 5.
Overfitting observed with ResNet18 CNN as sub-classifier for G3 grade classification in ‘Val1’ training set.
Table 2.
Architecture of the proposed CNN model for binary classification (one-vs-all).
Fig 6.
Training loss curves for the proposed CNN architecture as sub-classifier for G3 grade classification in ‘Val1’ training set after hyperparameter optimization.
Table 3.
Effect of L2 regularization parameter on F1-score (validation set).
Table 4.
Comparison of the proposed ensemble classifier on SICAPv2 dataset against reference works.
Fig 7.
Precision-Recall and F1-Score curves for the sub-classifiers in the ensemble detector on SICAPv2 ‘test’ set (a) G3 (b) G4 (c) G5.
Fig 8.
Precision-Recall and F1-Score curves for the sub-classifiers in the ensemble detector on SICAPv2 ‘val1’ set (a) G3 (b) G4 (c) G5.
Fig 9.
Precision-Recall and F1-Score curves for the sub-classifiers in the ensemble detector on SICAPv2 ‘val2’ set (a) G3 (b) G4 (c) G5.
Fig 10.
Precision-Recall and F1-Score curves for the sub-classifiers in the ensemble detector on SICAPv2 ‘val3’ set (a) G3 (b) G4 (c) G5.
Fig 11.
Precision-Recall and F1-Score curves for the sub-classifiers in the ensemble detector on SICAPv2 ‘val4’ set (a) G3 (b) G4 (c) G5.
Fig 12.
Grad-CAM visualization on example 1 a) input patch b) label mask c) G3 sub-classifier heat map d) G4 sub-classifier heat map e) G5 sub-classifier heat map.
Fig 13.
Grad-CAM visualization on example 2 a) input patch b) label mask c) G3 sub-classifier heat map d) G4 sub-classifier heat map e) G5 sub-classifier heat map.
Fig 14.
Grad-CAM visualization on example 3 a) input patch b) label mask c) G3 sub-classifier heat map d) G4 sub-classifier heat map e) G5 sub-classifier heat map.
Fig 15.
Activation maps on a WSI example a) Input image b) G3 c) G4 d) G5.