Table 3.
Literature review of publications for the prostate and IL segmentation
| Publication | Task | Method | Result | Evaluation |
|---|---|---|---|---|
| Tian et al17 | Prostate segmentation | Graph cut | DSC = 87.0% ± 3.2% | MICCAI 2012 Promise12 challenge |
| Mahapatra and Buhmann18 | Prostate segmentation | Super pixel + random forests + graph cut | DSC = 0.81 | MICCAI 2012 Promise12 challenge |
| Guo et al19 | Prostate segmentation | Stacked sparse auto-encoder + deformable segmentation | DSC = 0.871 ± 0.042 | 66 T2WIs |
| Milletari et al20 | Prostate segmentation | V-Net + dice-based loss | DSC = 0.869 ± 0.033 | Trained with 50 MRI scans Test with 30 MRI scans |
| Zhu et al21 | Prostate segmentation | Deeply supervised CNN | DSC = 0.885 | Trained with 77 patients Tested with 4 patients |
| Yu et al14 | Prostate segmentation | Volumetric convolutional neural network | DSC = 89.43% | MICCAI 2012 Promise12 challenge |
| Toth and Madabhushi22 | Prostate segmentation | Landmark-free AAM | DSC = 88% ± 5% | Tested with 108 studies |
| Liao et al23 | Prostate segmentation | Stacked independent subspace analysis + sparse label | DSC = 86.7% ± 2.2% | 30 T2WIs |
| Vincent et al24 | Prostate segmentation | AAM | DSC = 0.88 ± 0.03 | MICCAI 2012 Promise12 challenge |
| Klein et al25 | Prostate segmentation | Atlas matching | Median DSC varied between 0.85 and 0.88 | Leave-one-out test with 50 clinical scans |
| Li et al26 | Prostate segmentation | RW | DSC = 80.7% ± 5.1% | 30 MR volumes |
| Kohl et al27 | IL segmentation | Adversarial networks | DSC = 0.41 ± 0.28Sens. = 0.55 ± 0.36 Spec. = 0.98 ± 0.14 |
Four-fold cross-validation on 55 patients with aggressive tumor lesions |
| Cameron et al28 | IL detection | Morphology, asymmetry, physiology and size model | Accuracy (Acc.) = 87% ± 1% Sens. = 86% ± 3% Spec. = 88% ± 1% |
13 patients |
| Chung et al29 | IL segmentation | Radiomics-driven CRF | Sens. = 71.47% Spec.= 91.93% Acc. = 91.17% DSC = 39.13% |
20 patients |
| Artan et al30 | IL segmentation | Cost-sensitive support vector machine + CRF | Sens. = 0.84 ± 0.19 Spec. = 0.48 ± 0.22 DSC = 0.35 ± 0.18 |
21 patients |
| Artan et al31 | IL localization | RW | Sens. = 0.51 Jakkard = 0.44 |
10 patients |
| Artan et al32 | IL segmentation | RW | Sens. = 0.62 ± 0.23 Spec. = 0.89 ± 0.10 DSC = 0.57 ± 0.21 |
16 patients with lesions in peripheral zone only |
| Ozer et al33 | IL segmentation | Relevance vector machine | Spec. = 0.78 Sens. = 0.74 DSC = 0.48 |
20 patients |
| Artan et al34 | IL segmentation | Cost-sensitive CRF | Sens. = 0.73 ± 0.25 Spec. = 0.75 ± 0.13 Acc. = 0.71 ± 0.18 DSC = 0.45 ± 0.28 |
10 patients with lesions in peripheral zone only |
| Liu et al35 | IL segmentation | Fuzzy Markov random fields | Spec. = 89.58% Sens. = 87.50% Acc. = 89.38% DSC = 0.6222 |
11 patients |
Abbreviations: AAM = active appearance models; CNN = convolutional neural network; CRF = conditional random field; DSC = dice similarity coefficient; IL = intraprostatic lesions; MRI = magnetic resonance imaging; RW = random walker; T2WIs = T2-weighted images; Sens. = sensitivity; Spec = specificity.