Automated Classification of Circulating Tumor Cells and the Impact of Interobsever Variability on Classifier Training and Performance

Table 2

The performance of the classifiers when trained on two folds from probGT and GT in different combinations. It is cyclically tested on one GT fold that was not used for training.

	Training on two GT folds	Training on one GT fold and one probGT fold	Training on two probGT folds	Entire data set (across observers)
RF	Acc: 0.98 ± 0.00	Acc: 0.96 ± 0.01	Acc: 0.94 ± 0.02	Acc: 0.86 ± 0.04
	Pre: 0.98 ± 0.00	Pre: 0.94 ± 0.02	Pre: 0.89 ± 0.05	Pre: 0.83 ± 0.06
	Rec: 0.98 ± 0.00	Rec: 0.98 ± 0.01	Rec: 0.97 ± 0.03	Rec: 0.88 ± 0.08
SVM	Acc: 0.96 ± 0.00	Acc: 0.92 ± 0.01	Acc: 0.81 ± 0.04	Acc: 0.86 ± 0.03
	Pre: 0.95 ± 0.00	Pre: 0.88 ± 0.02	Pre: 0.75 ± 0.08	Pre: 0.85 ± 0.05
	Rec: 0.96 ± 0.00	Rec: 0.95 ± 0.01	Rec: 0.84 ± 0.07	Rec: 0.85 ± 0.08