Table 3.

Performances of various models in discriminating between stage T1 nasopharyngeal carcinoma (NPC) and benign hyperplasia in T2-weighted, fat-suppressed MRI. The accuracy, sensitivity, and specificity were evaluated using the threshold determined by the point maximizing the Youden index in the corresponding fold. As logistic regression returned binary results in our implementation, its threshold is effectively ≥1 for NPC.

Method	AUC	Accuracy	Sensitivity	Specificity	Threshold (≥)
SVR	0.85 ± 0.04	80.3% ± 1.8%	79.6% ± 9.7%	80.8% ± 10.0%	0.51 ± 0.10
Logistic Regression	0.77 ± 0.01	77.2% ± 1.2%	74.2% ± 3.9%	80.1% ± 3.3%	1
RF	0.82 ± 0.02	76.7% ± 1.4%	70.1% ± 8.4%	83.2% ± 7.4%	0.54 ± 0.10
Perceptron	0.84 ± 0.03	79.4% ± 1.6%	73.8% ± 4.8%	85.1% ± 4.4%	0.51 ± 0.06
kNN	0.82 ± 0.03	77.2% ± 1.4%	71.5% ± 7.9%	82.9% ± 9.2%	0.55 ± 0.10

The data presented are the average across the five folds ± standard error. Thresholds indicate the lower inclusive bound for a positive indication of NPC. SVR = support vector regression, RF = random forest, kNN = k-nearest neighbors.