TABLE 3.
Mean values and (standard deviations) for the Dice values, area under the ROC curve (AUC), volume errors, and absolute volume errors of each tissue outcome prediction method.
| Total | IA patients | IV patients | ||
| Dice | Tmax thresholding | 0.249 (0.214) | 0.217 (0.204) | 0.324 (0.220) |
| Random decision forest | 0.262 (0.213) | 0.206 (0.188) | 0.300 (0.211) | |
| Param-UNet | 0.287 (0.229) | 0.252 (0.217) | 0.369 (0.240) | |
| RC-Simple | 0.276 (0.232) | 0.259 (0.220) | 0.314 (0.244) | |
| RC-Causal | 0.286 (0.228) | 0.260 (0.219) | 0.346 (0.240) | |
| CTC-Causal | 0.296 (0.234) | 0.264 (0.211) | 0.384 (0.264) | |
| ROC-AUC | Tmax thresholding | 0.693 (0.320) | 0.690 (0.301) | 0.698 (0.362) |
| Random decision forest | 0.740 (0.125) | 0.741 (0.120) | 0.737 (0.140) | |
| Param-UNet | 0.773 (0.146) | 0.781 (0.147) | 0.768 (0.144) | |
| RC-Simple | 0.764 (0.184) | 0.770 (0.186) | 0.767 (0.182) | |
| RC-Causal | 0.768 (0.153) | 0.783 (0.147) | 0.747 (0.167) | |
| CTC-Causal | 0.791 (0.142) | 0.786 (0.133) | 0.802 (0.159) | |
| Volume error (ml) | Tmax thresholding | 83.8 (95.1) | 93.2 (85.1) | 61.42 (113.54) |
| Random decision forest | 48.2 (98.5) | 48.8 (83.5) | 46.5 (128.5) | |
| Param-UNet | 53.6 (100.0) | 53.8 (88.1) | 53.1 (124.9) | |
| RC-Simple | 66.4 (141.8) | 48.1 (82.1) | 109.9 (223.5) | |
| RC-Causal | 58.8 (122.4) | 36.5 (85.5) | 111.8 (172.4) | |
| CTC-Causal | 48.5 (93.6) | 50.1 (89.0) | 44.7 (104.8) | |
| Abs. volume error (ml) | Tmax thresholding | 107.8 (66.4) | 108.6 (64.1) | 106.0 (72.4) |
| Random decision forest | 86.6 (66.9) | 76.4 (59.0) | 110.9 (78.4) | |
| Param-UNet | 89.1 (70.1) | 81.2 (63.6) | 107.8 (81.3) | |
| RC-Simple | 106.3 (114.7) | 72.4 (61.5) | 186.8 (163.1) | |
| RC-Causal | 97.1 (94.8) | 68.4 (62.7) | 165.2 (120.8) | |
| CTC-Causal | 77.4 (71.5) | 76.7 (67.1) | 78.8 (81.7) |
IA indicates treatment with intra-arterial mechanical thrombectomy with or without thrombolysis, while IV indicates treatment with thrombolysis only.
The values corresponding to the best model performance for each performance metric are shown in bold.
Abbreviated model names correspond to: Tmax thresholding, random decision forests (RDF), deep learning from perfusion parameter maps (Param-UNet), deep learning from deconvolved residual curves with convolutional (RC-Simple) or causal convolutional (RC-Causal) feature extraction, and deep learning from source concentration-time curves (CTC-Causal).