Table 2.
Accuracy, macro-averaged F1 and area under the curve analysis of trained networks for multiclass classification (glioblastoma, brain metastasis, and meningioma) and binary classification (glioblastoma and brain metastasis).
| Network | Accuracy (%) | F1/AUC | Rate of diagnostic images (%) |
|---|---|---|---|
| ResNet18 | |||
| Test total | 85.8 | 92.3% (F1) | 32.3 |
| Test confidence > 0.999 | 93.6 | ||
| ResNet18 filtered | |||
| Test total | 87.3 | 93.2% (F1) | 36.3 |
| Test confidence > 0.999 | 98.6 | ||
| InceptionNet | |||
| Test total | 82.9 | 90.6% (F1) | 17 |
| Test confidence > 0.999 | 91.1 | ||
| ResNet18 binary classification | |||
| Test total | 90.9 | 0.92 (AUC) | 35.6 |
| Test confidence > 0.999 | 100 |
ResNet filtered contained manually selected data free of substantial artifacts. The ratio of images rated with an output level of 0.999 or higher and the amount of total images are indicated as diagnostic images. Images were obtained with CLE following topical ex vivo staining with fluorescein dye. Macro-averaged F1 was calculated using the following equation: 2*(precisionm*recallm)/(precisionm+recallm).