Table 1.
Comparison of the performance of different deep learning-based segmentation methods.
| Publication | Type of images | Proposed methods | Comparison baseline | ||
|---|---|---|---|---|---|
| Method | Results | Method | Results | ||
| Zhou et al. [14] | Multiple MRI | DNN | average = 0.864 (average of SEN, SPE and PRE) | Manifold learning | Average = 0.849 |
|
| |||||
| Zikic et al. [19] | BRAST 2013 | CNN | HGG (complete): ACC = 0.837 ± 0.094 | RF | HGG: ACC = 0.763 ± 0.124 |
|
| |||||
| Lyksborg et al. [20] | Multimodal MRI | CNN | Dice = 0.810, PPV = 0.833, SEN = 0.825 | Axially trained 2D network | Dice = 0.744, PPV = 0.732, SEN = 0.811 |
|
| |||||
| Dvořák and Menze [23] | BRATS 2014 | CNN | HGG (complete): Dice = 0.83 ± 0.13 | — | — |
|
| |||||
| Pereira et al. [24] | BRATS 2015 | CNN | LGG (complete): DSC = 0.86, PPV = 0.86, SEN = 0.88 HGG (complete): DSC = 0.87, PPV = 0.89, SEN = 0.86 Combined: DSC = 0.87, PPV = 0.89, SEN = 0.86 |
— | — |
|
| |||||
| Pereira et al. [25] | BRATS 2013 | CNN | DSC = 0.88, PPV = 0.88, SEN = 0.89 | Tumor growth model + tumor shape prior + EM | DSC = 0.88, PPV = 0.92, SEN = 0.84 |
|
| |||||
| Havaei et al. [27] | BRAST 2013 | INPUTCASCADECNN | Dice = 0.88, SPE = 0.89, SEN = 0.87 | RF | Dice = 0.87, SPE = 0.85, SEN = 0.89 |
|
| |||||
| Kamnitsas et al. [29] | BRATS 2015 | Multiscale 3D CNN + CRF | DSC = 0.849, PREC = 0.853, SEN = 0.877 |
— | — |
|
| |||||
| Yi et al. [32] | BRATS 2015 | 3D fully CNN | ACC = 0.89 | GLISTR algorithm | ACC = 0.88 |
|
| |||||
| Casamitjana et al. [33] | BRATS 2015 | Three different 3D fully connected CNNs | ACC = 0.9969/0.9971/0.9971 | — | — |
|
| |||||
| Zhao et al. [36] | BRATS 2013 | 3D fully CNN + CRF | Dice = 0.87, PPV = 0.92, SEN = 0.83 |
CNN | Dice = 0.88, PPV = 0.88, SEN = 0.89 |
|
| |||||
| Alex et al. [38] | BRATS 2013/2015 | SDAE | ACC = 0.85 ± 0.04/0.73 ± 0.25 | — | — |
|
| |||||
| Ibragimov et al. [39] | CT, MR and PET images | CNN | Dice = 0.818 | — | — |
Notes. BRAST = multimodal brain tumor segmentation dataset, including four MRI sequences (T1W, T1-postcontrast (T1c), T2W, and FLAIR); CNN = convolutional neural networks; HGG = high-grade gliomas; ACC = accuracy; RF = random forests; DNN = deep neural network; Average = the average values of sensitivity, specificity, and precision; LGG = low-grade gliomas; PPV = positive predictive value; SEN = sensitivity; DSC = dice similarity coefficient; INPUTCASCADECNN = cascaded architecture using input concatenation; EM = expectation maximization algorithm; SPE = specificity; PREC = precision; GLISRT (glioma image segmentation and registration); CRF = conditional random fields; SDAE = stacked denoising autoencoder.