Table 2.
Image level core estimation using our approach (with CTA input), Rapid.Ai (CTP input) and expert adjudication with CTA input (CTA-ASPECTS).
| RAPID.ai on CTP |
DeepSymNet-v3 on CTA |
Expert on CTA (CTA-ASPECTS) |
|||||||
|---|---|---|---|---|---|---|---|---|---|
| AuRoC | AuPrC | Sens/Spec | AuRoC | AuPrC | Sens/Spec | AuRoC | AuPrC | Sens/Spec | |
| Ischemic Core Threshold | |||||||||
| 15 ml | 0.77 | 0.89 | 0.56/0.90 | 0.83 | 0.93 | 0.62/0.95 | 0.76 | 0.88 | 0.74/0.71 |
| 30 ml | 0.81 | 0.82 | 0.62/0.89 | 0.85 | 0.83 | 0.91/0.65 | 0.82 | 0.79 | 0.88/0.65 |
| 50 ml | 0.82 | 0.78 | 0.71/0.91 | 0.87 | 0.78 | 0.79/0.87 | 0.81 | 0.65 | 0.92/0.55 |
| 70 ml | 0.86 | 0.76 | 0.82/0.94 | 0.90 | 0.69 | 0.88/0.85 | 0.78 | 0.50 | 0.94/0.50 |
The external test set with valid CTP (n = 71) is used and the final ground truth used is the average between the two reader segmentations on DWI. AuRoC: Area under the ROC curve; AuPrC: Area under the precision recall curve; Sens/Spec: Sensitivity/Specificity. In bold, the best performing method out of the three compared according to AuRoC or AuPrC. For each row and metric, the best performing method is highlighted in bold.