Table 2.
Diagnostic performance of traditional machine learning, deep learning, and human readers.
| Data | Classifier | Cross validation | External validation | |||
|---|---|---|---|---|---|---|
| Mean AUC | AUC | Sensitivity (%) | Specificity (%) | Accuracy (%) | ||
| Traditional machine learning | ||||||
| CE mask | AdaBoost with LASSO | 0.870 | 0.858 (0.787–0.926) | 68.0 (53.3–80.5) | 93.8 (79.2–99.2) | 78.0 (67.5–86.4) |
| L-SVM with Tree-based selection | 0.875 | 0.833 (0.755–0.904) | 62.0 (47.2–75.4) | 93.8 (79.2–99.2) | 74.4 (63.6–83.4) | |
| LDA with LASSO | 0.863 | 0.818 (0.737–0.891) | 64.0 (49.2–77.1) | 87.5 (71.0–96.5) | 73.2 (62.2–82.4) | |
| PT mask | AdaBoost with Tree-based selection | 0.816 | 0.773 (0.668–0.870) | 94.0 (83.5–98.8) | 34.4 (18.6–53.2) | 70.7 (59.7–80.3) |
| L-SVM with RFE | 0.830 | 0.803 (0.718–0.879) | 86.0 (75.5–96.5) | 65.6 (46.8–81.4) | 78.0 (67.2–88.8) | |
| LDA with MI | 0.818 | 0.787 (0.695–0.870) | 94.0 (83.5–98.8)) | 50.0 (31.9–68.1 | 76.8 (66.2–85.4) | |
| Combined mask | AdaBoost with Tree-based selection | 0.926 | 0.890 (0.823–0.947) | 80.0 (62.3–90.0) | 87.5 (71.0–94.5) | 82.9 (73.0–90.3) |
| L-SVM with RFE | 0.932 | 0.886 (0.798–0.927) | 80.0 (62.3–90.0) | 84.4 (67.2–94.7) | 81.7 (71.6–89.4) | |
| LDA with LASSO | 0.945 | 0.899 (0.839–0.951) | 84.0 (70.9–92.8) | 78.1 (70.9–90.7) | 81.7 (71.6–89.4) | |
| Deep learning | ||||||
| CE mask | DNN | 0.887 | 0.887 (0.812–0.951) | 62.5 (45.7–79.3) | 96.0 (90.6–100) | 82.9 (74.8–91.1) |
| PT mask | DNN | 0.865 | 0.825 (0.722–0.887) | 75.0 (60.2–90.1) | 82.0 (71.4–92.6) | 79.3 (70.5–88.0) |
| Combined mask | DNN | 0.986 | 0.956 (0.918–0.990) | 90.6 (80.5–100) | 88.0 (79.0–97.0) | 89.0 (82.3–95.8) |
| Human reading | ||||||
| Images | Reader 1 | 0.774 (0.685–0.852) | 97.0 (91.1–100) | 50.0 (36.1–63.9) | 68.7 (58.7–78.7) | |
| Images | Reader 2 | 0.904 (0.852–0.951) | 81.8 (68.7–95.0) | 78.0 (66.5–89.5) | 79.5 (70.8–88.2) | |
Values in parentheses are 95% confidence intervals.
AUC area under the receiver operating characteristic curve, CE contrast-enhancing, PT peritumoral T2 hyperintense, AdaBoost adaptive boosting, L-SVM linear support vector machine, LDA linear discriminant analysis, RFE recursive feature elimination, DNN deep neural net.