Table 4 Comparison of the existing state-of-the-arts on MICCAI MSSEG 2016 Training Data65,68. (TPR = TPF; PPV = 1 − FPF).
From: Radius-optimized efficient template matching for lesion detection from brain images
Group | Method | MRI sequence | Method type | Performance | ||
---|---|---|---|---|---|---|
DICE | TPR | PPV | ||||
Mahbod et al.59 | Multilayer perceptron with morphology-based filtering | 3D FLAIR | Sup. Seg. | 10.2–84.0 | – | – |
Vera-Olmos et al.60 | RF classifier and MRF based post-processing | T1, FLAIR | Sup. Seg. | 63.8 | 68.3 | – |
Salehi et al.61 | 3D Fully convolutional network with Tversky loss | T1, T2, FLAIR | Sup. Seg. | 56.4 | 56.8 | – |
Hashemi et al.62 | Patch-wise 3D fully convolutional DenseNet architecture | T1, T1-GADO, FLAIR, PD, T2 | Sup. Seg. | 69.9 | 78.5 | – |
Coupe et al.63 | Rotationally-invariant NLM and patch-wise NLM denoising filter | T1, FLAIR | Sup. Seg. | 72.5 | – | – |
Chen et al.64 | Hybrid feature network based on DenseNet architecture | T1, T2, FLAIR | Sup. Seg. | 66.5 | 61.3 | – |
Kamraoui et al.65 | DeepLesionBrain (DLB) | T1, FLAIR | Sup. Seg. | 63.9 | 60.8 | 76.8 |
DLB with hierarchical specialization learning | T1, FLAIR | 66.9 | 67.1 | 72.8 | ||
Valverde et al.66 | 3D cascaded CNN with \(11\times 11\times 11\) patch | T1, FLAIR | Sup. Seg. | 44.2 | 42.3 | 61.4 |
Zhang et al.67 | 2.5D densely connected fully convolutional network | T1, FLAIR | Sup. Seg. | 66.4 | 65.8 | 74.1 |
McKinley et al.68 | 3D-2D CNN (DeepSCAN) architecture | T1, T2, FLAIR | Sup. Seg. | 75.7 | – | – |
Valverde et al.44 | Cascaded 3D CNN | T1, T2, FLAIR | Sup. Seg. | 58.7 | – | – |
Isensee et al.69 | 2D, 3D, cascade of two 3D U-Net | T1, T2, FLAIR | Sup. Seg. | 74.5 | – | – |
Beaumont et al.58 | Multimodal graph cut, EM, and post-processing | T1, T2, FLAIR | UnSup. Seg. | 57.0 | – | – |
Beaumont et al.70 | Voxel-wise comparison (GMM and EM) and post-processing | FLAIR | UnSup. Seg. | 50.5 | – | – |
T2 \(\cup\) FLAIR | 42.4 | – | – | |||
T2 \(\cap\) FLAIR | 43.9 | – | – | |||
FLAIR and T2 \(\cap\) FLAIR | 56.6 | – | – | |||
Beaumont et al.70 | Voxel-wise comparison without post-processing | FLAIR | UnSup. Seg. | 42.3 | – | – |
T2 \(\cup\) FLAIR | 29.7 | – | – | |||
T2 \(\cap\) FLAIR | 24.7 | – | – | |||
Knight and Khademi71 | Fuzzy classification, thresholding, post-processing | FLAIR | UnSup. Seg. | 60.0 | 53.0 | 80.0 |
Baseline method | FFT-based template matching in \(\mathcal{O}\left({a}_{max}N\mathrm{log}N\right)\) | FLAIR | UnSup. Detection | 11.0 | 20.0 | 11.8 |
8.8 | 40.0 | 6.1 | ||||
6.3 | 60.0 | 3.8 | ||||
4.4 | 80.0 | 2.4 | ||||
Proposed | Template matching in \(\mathcal{O}\left({a}_{max}N\right)\) | FLAIR | UnSup. Detection | 11.2 | 20.0 | 12.2 |
9.5 | 40.0 | 6.8 | ||||
6.7 | 60.0 | 4.0 | ||||
4.6 | 80.0 | 2.5 | ||||
Proposed | Template matching in \(\mathcal{O}\left(N\right)\) | FLAIR | UnSup. Detection | 13.5 | 20.0 | 20.2 |
12.0 | 40.0 | 9.6 | ||||
8.5 | 60.0 | 5.4 | ||||
6.2 | 80.0 | 3.5 |