Table 1.
Cross-validation results for the different classifier candidates.
| method | Freezer N = 382/143 | Precultured N = 1946/294 | ||||||
|---|---|---|---|---|---|---|---|---|
| AUC | ACC | F1 | SensSpec | AUC | ACC | F1 | SensSpec | |
| Logistic regression | 85.1 | 77.5 | 65.3 | 77.6 | 89.1 | 81.7 | 53.9 | 81.6 |
| Random forest | 87.8 | 80.4 | 69.1 | 80.4 | 86.3 | 78.8 | 49.5 | 78.9 |
| SVM | 92.0 | 82.7 | 72.2 | 82.5 | 88.2 | 79.9 | 51.1 | 79.9 |
| Bundle | 88.4 | 78.3 | 65.9 | 76.9 | 88.4 | 80.8 | 52.6 | 81.0 |
The results are displayed separately for the two culturing methods that are relevant to the task. The metrics are AUC = area under the ROC curve, ACC = Accuracy, F1 = the harmonic mean of precision and recall, and SensSpec = the sensitivity and specificity of the classifier. Images of cells cultured as cell lines do not contain any fibroblasts and are classified perfectly by each of the tested classifiers. The best variation of each method was chosen based on the area under the AUC. Reference class for the metrics is cancer.