Skip to main content

Table 1 Overview of previous studies applying neural networks for the detection of AD and MCI

From: Improving 3D convolutional neural network comprehensibility via interactive visualization of relevance maps: evaluation in Alzheimer’s disease

Study (chronologic order) Data type Sample Algorithm Performance   Addressed model comprehensibility
AD MCI
c/nc
CN   Groups Accuracy Balanced accuracy AUC  
Suk et al. [17] MRI GM and FDG-PET 93 76/128 101 RBM on class discriminative patches selected by statistical significance tests AD/CN
MCI/CN
MCIc/MCInc
95.4%
85.7%
74.6%
94.9%
80.6%
71.6%
0.988
0.881
0.747
Visualization of selected features (image patches) and RBM model weights projected on MRI scan
Li et al. [18] MRI and FDG-PET 51 43/56 52 RBM for feature learning, SVM for classification AD/CN
MCI/CN
MCIc/MCInc
91.4%
77.4%
57.4%
   No
Ortiz et al. [19] MRI GM and FDG-PET 70 39/64 68 RBM for feature learning, SVM for classification AD/CN
MCIc/CN
MCIc/MCInc
90%
83%
78%
  0.95
0.95
0.82
Visualization of SVM model weights projected on MRI scan
Aderghal et al. [20] MRI and DTI 188 339 228 CNN for hippocampus region of interest only AD/CN
MCI/CN
92.5%
80.0%
92.5%
82.9%
  No
Liu et al. [21] FDG-PET 93 146 100 CNN and RNN AD/CN
MCI/CN
91.2%
78.9%
  0.953
0.839
Visualization of most contributing brain areas obtained from occlusion sensitivity analysis
Liu et al. [22] MRI 199 229 CNN on landmarks selected by statistical significance tests AD/CN
MCIc/CN
90.6%   0.957 Visualization of top 50 anatomical landmarks used as input for the CNN
Lin et al. [23] MRI 188 169/193 229 CNN AD/CN
MCIc/MCInc
88.8%
79.9%
  0.861 No
Böhle et al. [14] MRI 211 169 CNN AD/CN 88.0%    Visualization of LRP relevance and guided backpropagation maps, comparison of LRP relevance scores by group and brain region
Li et al. [24] MRI Training 192
Test 225
383
479
228
639
CNN for hippocampus only AD/CN
MCIc/MCInc
92.9%   0.958
0.891
Visualization of most contributing hippocampus areas obtained from CNN class activation mapping
Dyrba et al. [11] MRI 189 219 254 CNN for coronal slices covering hippocampus AD/CN
MCI/CN
   0.93
0.75
Visualization of LRP and other methods’ relevance maps and comparison by diagnostic group
Lian et al. [25] MRI Training 199
Test 159
167/226
38/239
229
200
CNN AD/CN
MCIc/MCInc
90.3%
80.9%
  0.951
0.781
Visualization of most contributing image areas obtained from CNN class activation mapping
Qiu et al. [26] MRI Training 188
Test1 62
Test2 29
Test3 209



229
320
73
356
FCN AD/CN1
AD/CN2
AD/CN3
87.0%
76.6%
81.8%
  0.870
0.892
0.881
Visualization of most contributing brain areas obtained from occlusion sensitivity analysis
Wen et al. [27] MRI Training 336
Test1 76
Test2 78
295/298
20/13
330
429
76
CNN AD/CN1
MCIc/MCInc1
AD/CN2
  86%
50%
70%
  No
Thibeau-Sutre et al. [8] MRI Training 336
Test 76

330
429
CNN AD/CN   90%   Visualization of most contributing brain areas obtained from occlusion sensitivity analysis
Jo et al. [28] Tau-PET 66 66 CNN AD/CN 90.8%    Visualization of LRP relevance maps, visualization of most contributing brain areas obtained from occlusion sensitivity analysis
  1. Empty cells in the performance columns indicate that the respective values were not reported
  2. AD Alzheimer’s dementia, MCI mild cognitive impairment, MCIc MCI converted to dementia, MCInc non-converter/stable MCI, CN cognitively normal controls, DTI diffusion tensor imaging, FCN fully connected network, RBM restricted Boltzmann machine, RNN recurrent neural network, CNN convolutional neural network, MRI T1-weighted magnetic resonance imaging, GM gray matter volume, FDG-PET glucose metabolism derived from fluorodeoxyglucose positron emission tomography