Table 4.
Comparison of the performance of the proposed CMB detection method with existing conventional machine learning (ML) and deep learning (DL) methods.
| Methods | Datasets | Performance | |||
|---|---|---|---|---|---|
| Sequence(s) (# test subjects) | Total # CMBs | Cl. TPR | FPavg | Cl. Prec | |
| ML methods | |||||
| Bian et al. (2013) | SWI (10) | 304 | 86.5% | 44.9 | |
| Fazlollahi et al. (2014) | SWI (41) | 103 | 92% | FPavgCMB - 6.7 FPavgnCMB - 16.8 | |
| Fazlollahi et al. (2015) | SWI (66) | 231 | 87% | FPavgD - 10.28, FPavgP+D - 27.8 | |
| Ghafaryasl et al. (2012) | T2*-GRE + PD (81) | 183 | 91% | 4.1 | |
| Dou et al. (2015) | SWI (19) | 161 | 80% | 7.7 | 49% |
| Chesebro et al. (2021) | T2*-GRE, SWI (78) | 64 | 95% | 9.7 (SWI), 17.1 (T2*-GRE) | 11% (SWI),7% (T2*-GRE) |
| DL methods | |||||
| Chen et al. (2015) | SWI (5) | 55 | 89% | 6.4 | 56% |
| Dou et al. (2016) | SWI (50) | 117 | 93% | 2.74 | 44% |
| Liu et al. (2019b) | Phase + SWI (41) | 168 | 96% | 1.8 | (5-fold CV) |
| Al-Masni et al. (2020) | Phase + SWI (72) | 188 | 94.3% | 1.4 | 61.9% |
| Rashid et al. (2021) (Leave-one-out validation) | QSM + SWI + T2w (24) | ~172 | 89% | 49% | |
| Proposed method | UKBB - SWI (78) | 186 | 93% | 1.5 | 59% |
| OXVASC - T2*-GRE (74) | 366 | 90% | 0.9 | 84% | |
Cl.TPR, cluster-wise TPR; FPavg, average false positives per image/subject; Cl.Prec, cluster-wise precision; FPavgCMB, FPavg for CMB subjects; FPavgnCMB, FPavg for non-CMB subjects; FPavgD - FPavg for “definite” CMB subjects; FPavgP+D, FPavg for “definite and possible” CMB subjects.