Table 3.
Evaluation of the generalizability of the proposed method - trained on the UKBB SWI data and evaluated on the UKBB QSM, TICH2, SHK, and OXVASC datasets: performance metrics at candidate detection, discrimination and post-processing steps.
| Datasets | Steps | Cl. TPR | FPavg | Cl. prec |
|---|---|---|---|---|
| UKBB (QSM) | C. det. | 0.99 | 138.0 | 0.02 |
| C. disc. | 0.91 | 40.3 | 0.04 | |
| Postproc. | 0.90 | 1.8 | 0.44 | |
| TICH2 (SWI) | C. det. | 0.88 | 289.3 | 0.02 |
| C. disc. | 0.83 | 42.8 | 0.10 | |
| Postproc. | 0.82 | 3.1 | 0.62 | |
| SHK (SWI) | C. det. | 0.98 | 254.7 | 0.01 |
| C. disc. | 0.94 | 43.6 | 0.09 | |
| Postproc. | 0.87 | 0.5 | 0.89 | |
| OXVASC (T2*-GRE) | C. det. | 0.88 | 147.1 | 0.03 |
| C. disc. | 0.85 | 53.7 | 0.07 | |
| Postproc. | 0.81 | 2.0 | 0.71 |
Cl. TPR and Cl. prec indicate cluster-wise TPR and cluster-wise precision, respectively. C. det, candidate detection; C. disc, candidate discrimination. FPavg at cluster-wise TPR value of 95% could not be provided since specific values of thresholds were applied.