Skip to main content
. 2023 Mar 25;90:104540. doi: 10.1016/j.ebiom.2023.104540

Table 1.

Literature review.

Author Samples Methods Results Gaps addressed in this report
Odusami et al. (2022)14 Healthy controls (CN) = 25
Subjective memory complaint (SCM) = 25
Early mild cognitive impairment (EMCI) = 25
Late MCI = 25
MCI = 13
Alzheimer's Disease (AD) = 25		 Concatenated extracted features from MRI acquisitions using ResNet18 and DenseNet121 98.86% accuracy, 98.94% precision, and 98.89% recall in multiclass classification No assessment of non-stereotypical samples
Trained and tested with data from one source only (ADNI)
No data augmentation
No dementia subtype characterisation (all patients had the AD spectrum)
No use of demographically matched sample in age, sex, and years of education to discard possible biases across groups inflating classification accuracy
Razzak et al. (2022)15 AD = 95
MCI = 138
CN = 146		 PartialNet Models averaged 99% accuracy No assessment of non-stereotypical samples
Image preprocessing not fully automatic
Trained and tested with data from one source only (ADNI)
No dementia subtype characterisation (all patients had the AD spectrum)
No model interpretability analysis
No use of demographically matched sample in age, sex, and years of education to discard possible biases across groups inflating classification accuracy
Di Benedetto et al.(2022)16 Database 1:
Behavioural-variant frontotemporal dementia (bvFTD) = 50
CN = 110
Database 2:
bvFTD = 29
CN = 24		 Compared 3D CNN, vision transformers, and logistic regression Up to 90% AUC No assessment of non-stereotypical samples
Preprocessed MRI using CAT12 toolbox
No data augmentation
No model interpretability analysis
No dementia subtype characterisation (all patients were bvFTD)
No use of demographically matched sample in age, sex, and years of education to discard possible biases across groups inflating classification accuracy
Hosseini-Asl et al. (2018)17 MCI = 70
AD = 70
CN = 70		 3D CNN autoencoder Up to 100% accuracy Trained and tested with data from one source only (ADNI)
Image preprocessing
No dementia subtype characterisation (all patients had the AD spectrum)
No model interpretability analysis
No use of demographically matched sample in age, sex, and years of education to discard possible biases across groups inflating classification accuracy
Qiu et al. (2020)18 Database 1:
CN = 229 AD = 118
Database 2:
CN = 320 AD = 62
Database 3:
CN = 73 AD = 62
Database 4:
CN = 356 AD = 209		 3D CNN AUC up to 0.996 No assessment of non-stereotypical samples
No data augmentation
Image preprocessing
No dementia subtype characterisation
All samples from high income country databases, no 1.5 Tesla testing samples
Two databases were not demographically matched sample in age, sex, and years of education to discard possible biases across groups inflating classification accuracy
Amini et al. (2021)19 Low MCI = 690
Mild MCI = 236
Moderate MCI = 67
Severe MCI = 7		 Quantum Matched-Filter Technique (QMFT) and CNN deep learning Up to 96.7% accuracy No assessment of non-stereotypical samples
Trained and tested with data from one source only (ADNI)
Image preprocessing
No dementia subtype characterisation (all patients had the AD spectrum)
No use of demographically matched sample in age, sex, and years of education to discard possible biases across groups inflating classification accuracy
Zhang et al. (2021)20 AD = 280
cMCI = 162 (MCI converters)
ncMCI = 251 (MCI non-converters)
CN = 275		 Attentional CNN Up to 97.35% accuracy No assessment of non-stereotypical samples
Trained and tested with data from one source only (ADNI)
Image preprocessing
No dementia subtype characterisation (all patients had the AD spectrum)
No model interpretability analysis
No use of demographically matched sample in age, sex, and years of education to discard possible biases across groups inflating classification accuracy
Feng et al. (2020)21 AD = 153
MCI = 157
CN = 159		 3D-CNN Up to 99.1% accuracy No assessment of non-stereotypical samples
Trained and tested with data from one source only (ADNI)
No dementia subtype characterisation (all patients had the AD spectrum)
No model interpretability analysis
No use of demographically matched sample in age, sex, and years of education to discard possible biases across groups inflating classification accuracy