Table 3.
Reconstruction (1. training phase) results: Average Dice values on the test data for the CAE trained on core (Sc) and core+penumbra (Scp) expert segmentations, as well as average Dice overlap of the follow-up expert lesion segmentation (Sl) with the reconstructed interpolation from shape space (Ri).
| Method | Method | # Parameters | Dice | ||
|---|---|---|---|---|---|
| Core/Penumbra | Lesion (Training phase) | of network | Core (Rc) | Core+Penumbra (Rcp) | Lesion (Ri) |
| Expert | Oracle CAE 10 h (1. phase) | 4.7·106 | 0.68 | 0.90 | 0.53 |
| Expert | CAE 10 h (1.) | 4.7·106 | 0.68 | 0.90 | 0.46 |
| Expert | CAE 24 h (1.) | 4.7·106 | 0.70 | 0.90 | 0.44 |
While the first row shows an oracle prediction overlap for the theoretically best-fit per case interpolation within our model (could be before or after the true time-to-treatment), the second row lists the result for our proposed approach using the actual time-to-treatment ground truth to predict the follow-up lesion from ground truth. The third row indicates that the effect of choosing a different normalization value from within the time range of the acute stroke phase can rather be neglected.