Table 2.
Comparison of the performance of deep learning-based classification methods.
| Publication | Type of images | Proposed methods | Comparison baseline | ||
|---|---|---|---|---|---|
| Method | Results | Method | Results | ||
| Reda et al. [40] | DW-MRI | SNCAE | ACC = 1, SEN = 1, SPE = 1 | K ∗ | ACC = 0.943, SEN = 0.943, SPE = 0.944 |
|
| |||||
| Reda et al. [41] | DW-MRI | SNCAE | ACC = 1, SEN = 1, SPE = 1, AUC ≈ 1 | K ∗ | ACC = 0.943, SEN = 0.962, SPE = 0.926, AUC = 0.93 |
|
| |||||
| Zhu et al. [42] | T2-weighted, DWI and ADC | SAE | SBE = 0.8990 ± 0.0423, SEN = 0.9151 ± 0.0253, SPE = 0.8847 ± 0.0389 | HOG features | SBE = 0.8814 ± 0.0534, SEN = 0.9191 ± 0.0296, SPE = 0.8696 ± 0.0563 |
|
| |||||
| Akkus et al. [43] | T1-postcontrast (T1C) and T2 | Multiscale CNN | ACC = 0.877, SEN = 0.933, SPE = 0.822 | — | — |
|
| |||||
| Pan et al. [44] | BRATS 2014 | CNN | SEN = 0.6667, SPE = 0.6667 | NN | SEN = 0.5677, SPE = 0.5677 |
|
| |||||
| Hirata et al. [45] | FDG PET | CNN | ACC = 0.88 | SUVmax | ACC = 0.80 |
|
| |||||
| Hirata et al. [46] | MET PET | CNN | ACC = 0.888 ± 0.055 | SUVmax | ACC = 0.66 |
|
| |||||
| Teramoto et al. [47] | PET/CT | CNN | SEN = 0.901, with 4.9 FPs/case | Active contour filter | SEN = 0.901, with 9.8 FPs/case |
|
| |||||
| Wang et al. [48] | FDG PET | CNN | ACC = 0.8564 ± 0.0809, SEN = 0.8353 ± 0.1385, SPE = 0.8775 ± 0.1030 AUC = 0.9086 ± 0.0865 | AdaBoost + D13 | ACC = 0.8505 ± 0.0897, SEN = 0.8565 ± 0.1346, SPE = 0.8445 ± 0.1261 AUC = 0.9143 ± 0.0751 |
|
| |||||
| Antropova et al. [51] | DCE-MRI | CNN ConvNet | AUC = 0.85 | — | — |
Notes. DW-MRI = diffusion-weighted magnetic resonance images; SNCAE = stacked nonnegativity-constrained autoencoders; ACC = accuracy; SEN = sensitivity; SPE = specificity; AUC = area under the receiver operating characteristic curve; K∗ = K-Star, a classifier implemented in Weka toolbox [59]; DWI = diffusion-weighted imaging; ADC = apparent diffusion coefficient; SAE = stacked autoencoder; SBE = section-based evaluation; HOG = histogram of oriented gradient; CNN = convolutional neural network; BRATS = multimodal brain tumor segmentation dataset, including four MRI sequences (T1W, T1-postcontrast, T2W, and FLAIR); NN = neural network; FDG = fluorodeoxyglucose; PET = positron emission tomography; SUVmax = maximum standardized uptake value; MET = 11C-methionine; CT = computed tomography; FP = false positive; AdaBoost = adaptive boosting; D13 = 13 diagnostic features.