[Retracted] An Exploration: Alzheimer’s Disease Classification Based on Convolutional Neural Network

Table 2

Characteristics of each of the study included in survey.

Ref no. (year)	Model	Data handling approach	Preprocessing techniques	Modality		Dataset	Cohort or subjects					Classification				Data augmentation	Transfer learning
				1	>1		AD	MCI		NC	Total	Binary			Multiclass (3-way or 4-way)
								cMCI	ncMCI			AD vs. HC	AD vs. MCI	MCI vs. HC
[2] (2017)	SVM	Slice based	Noise removal, linear normalization, image enhancement	MRI	—	ADNI	—	—	—	—	—	Sensitivity: 95.3% specificity:71.4%				—	No
	2D CNN						7	14		15	36	Sensitivity: 96% specificity:98%				—
	2D CNN						9	16		11	36	Sensitivity: 95% specificity:98%				—
[26] (2017)	2D CNN	Slice based	Image alignment, image normalization	MRI	—	ADNI	188	399		288	875	82.20%	62.50%	66%	—	Yes	No
[27] (2017)	2D CNN	Slice based	Geometric normalization for registration	MRI	—	ADNI	188	399		288	875	91%	—	—	—	Yes	No
[28] (2017)	2D CNN	Slice based	Gradwarp, B1 nonuniformity, N3	MRI	—	ADNI	47	—		34	81	93%	—	—	—	Yes	No
[29] (2018)	2D CNN	Slice based	Skull stripping, spatial normalization and smoothing	MRI	—	OASIS+community advertisements	28+70	—		98+0	196	97.65%	—	—	—	Yes	No
[11] (2018)	2D CNN	Slice based	—	MRI	—	OASIS	—	—		—	416	—	—	—	93.18%	Yes	No
[30] (2018)	2D CNN+RNN	Slice based	No segmentation and rigid registration	FDG-PET	—	ADNI	98	146		100	339	—	—	78.9%	—	Yes	No
[31] 2013	Sparse encoder+2D CNN	Patch based	Normalization using statistical parametric mapping (SPM)	sMRI	—	ADNI	200	411		232	843	93.8%	86.3%	83.3%	78.2%	Yes	Yes
[32] (2015)	Sparse encoder+2D CNN	Slice based	Normalization using statistical parametric mapping (SPM)	MRI	—	ADNI	755	755		755	2265	95.39%	82.24%	90.13%	85.53%	Yes	No
	Sparse encoder+3D CNN	Voxel based										98.85%	86.84%	92.11%	89.47%
[33] (2016)	2D CNN	Slice based	Motion correction, skull stripping, and spatial smoothing	rs-fMRI	—	ADNI	28	—		15	43	98.85%	—	—	—	Yes	No
[34] (2016)	2D CNN	Slice based	Skull striping, registration, spatial smoothing	MRI	—	ADNI	211	—		91	302	98.84%	—	—	—	Yes	No
[3] (2017)	2D CNN	Slice based	Skull stripping, spatial smoothing, registration using MNI	MRI	—	ADNI	211	—		91	302	98.84%	—	—	—	Yes	No
		Subject based										100%
		Slice based		rs-fMRI			52			92	144	99.9%
		Subject based										97.77%
[35] (2017)	2D CNN based	Slice based	Skull stripping and the GM segmentation	MRI	—	ADNI	33	22	49	45	149	—	—	—	4-way based	Yes	No
	GoogleNet														98.9%
	ResNet-18														98.01%
	ResNet-152														98.14%
[36] (2016)	3D CAE+3DCNN (3D ACNN)	Voxel based	No preprocessing	MRI	—	CADDementia MRI and validated on ADNI	70	70		70	210	97.60%	95%	90.80%	89.1% (3-way)	No	Yes
[37] (2016)	Deeply supervised adaptive 3D-CNN (DSA-3D CNN)	Voxel based	Used no preprocessing techniques	MRI	—	ADNI MRI and validated on CADDementia	70	70		70	210	99.30%	100.00%	94.20%	94.8% (3-way)	—	Yes
[38] (2017)	3D CNN+3D CAE	Voxel based	Skull stripping and cerebellum-removal (after an intensity inhomogeneity correction)	MRI		ADNI	199	—		229	428	88.31%	—	—	—	—	—
[39] (2017)	SAE+3DCNN	Patch based ()	Anterior commissure posterior commissure (AC-PC) correction, skull stripping, and cerebellum removal	—	MRI and PET	ADNI	145	192		172	509	93.14%	82.36%	89.47%	86.13%	Yes	Yes
												93.59%	82.92%	93.25	89.24%
												91.06%	83.75%	91.14%	87.53%
[40] (2017)	3D CNN	Patch based	Skull stripping, cerebellum removal, AC-PC correction	MRI	—	ADNI	145	—		172	317	80.62%	—	—	—	—	—
	SAE+3D CNN											85.24%
	3D CNN			PET	—							81.93%
	SAE+3D CNN											85.53%
	3D CNN			—	MRI and PET							84.72
	SAE+3D CNN											91.14%
[41] (2018)	3D CNN	ROI based	Skull stripping, coregistration, spatially normalized	—	MRI+DTI	ADNI	48	108		58	214	85%	75%	66%	—	No	—
												96.70%	80%	65.80%		Yes
[42] (2017)	3D CNN	Patch based	Correction of intensity inhomogeneity, skull stripping, and cerebellum removal	MRI	—	ADNI	199	—		299	498	87.15%	—	—	—	—	No
[43] (2019)	3D CNN	Slice based	Normalization using statistical parametric mapping and diffeomorphic anatomical registration exponentiated lie algebra (DARTEL)	sMRI		ADNI+non-ADNI (“Milan”)						99.2% (with ADNI) 98.2% (with ADNI and Milan)	—	—	—	Yes	Yes
[44] (2019)	3D CNN	ROI based	—	MRI	—	ADNI	647	326	441	731	2145	81.19%	—	—	—	—	No
				FDG-PET	—							89.11%
				—	MRI and FDG-PET							90.10%
[45] (2019)	3D CNN	ROI based	AC_PC, tissue intensity inhomogeneity, skull stripping and cerebellum removal, registration	—	MRI and PET	ADNI	93	76	128	100	397	94.82%	—	4.5—	—	—	No
[46] (2017)	3D CNN based on VGGNet and ResNet	Voxel based	Skull stripping, spatially normalized	MRI	—	ADNI	50	43	77	61	231	88% (VGGNet)	—	—	—	—	No
[47] (2018)	3D CNN (based on ResNet)	Voxel based	—	MRI	—	ADNI	345	450		574	1370	94%	—	90%	87%	—	No
[48] (2018)	3D CNN followed by 2D CNN	Voxel based	No preprocessing	—	MRI and PET	ADNI	93	—		100	193	89.64%	—	—	—	Yes	No
[10] (2018)	3D CNN followed by 2D CNN	Patch based	No segmentation and rigid registration	—	MRI and PET	ADNI	93	76	128	100	397	93.29%	—	—	—	Yes	Yes
[49] (2014)	3D CNN	Patch based	Intensity inhomogeneity, skull stripping and cerebellum removal	—	MRI and PET	ADNI	198	167	236	229	830	92.875	—	76.21%	—	Yes	No
[50] (2017)	2D CNN	Slice based	Motion correction, skull stripping and, intensity normalization	MRI	—	ADNI	300	300		300	900	—	—	—	91.85%	Yes	Yes
[51] (2017)	3D CNN	Patch based	Intensity normalization and coregistration	PET	—	ADNI	93	—		100	193	92.20%	—	—	—	Yes	No
[52] (2018)	CaffeNet	Slice based	Gradwarp, intensity inhomogeneity correction, and N3 histogram peak sharpening	MRI	—	ADNI	—	157	150	457	764	—	—	—	87.78%	Yes	Yes
	GoogleNet														83.23%
[53] (2018)	3D CNN	Voxel based	Registration, histogram matching	MRI+clinical assessment and genetic (APOe4)	—	ADNI	192			184	376	99%	—	—	—	—	No
[54] (2018)	2D CNN	Slice based	Spatially normalized, skull stripping	MRI	—	ADNI	150	129		112	391	95.91%	86.84%		89.76%	Yes	No
	3D CNN											96.81%	88.435		91.32%
[55] (2018)	En3DCNN	ROI based	Nonuniformity (NU) intensity correction, motion correction, Talairach space conversion	MRI	—	ADNI	347	—		417	764	93.90%	—	—	—	Yes	No
[56] (2018)	2D CNN	Slice based	Skull stripping	MRI	—	ADNI	347	806		537	1690	94.97%	91.98%	74.7%	—	Yes	Yes
[57] (2018)	2D CNN	Slice based	Gradient unwarping, nonparametric nonuniformed bias correction	sMRI	—	ADNI	336	542		785	1663	95.45%	93.88%	95.39%	—	—	—
[15] (2019)	2D CNN	Slice based	Skull stripping, motion correction, and NU intensity normalization	MRI	—	ADNI	50	50		50	150	99.14%	99.3%	99.22%	95.73%	Yes	Yes
[58] (2017)	Expedited CNN	Voxel based	—	sMRI	—	ADNI	—	400		229	629	—	—	88.8%	—	Yes	With LIDC
														90.6%			With OASIS
[12] (2017)	2D CNN based DenseNet-121-161-169	Slice based	—	MRI	—	OASIS		—		—	416	—	—	—	93.18%	No	Yes
[59] (2017)	VGG-16 (from scratch)	Slice based	—	sMRI	—	OASIS	100	—		100	200	74.13%	—	—	—	Yes	Yes
	VGG-16 (transfer learning)											92.3%
	Inception V4 (transfer learning)											96.25%
[60] (2019)	3D CNN	Voxel based	Skull stripping, bias field correction, volumetric and affine registration	MRI	—	ADNI	—	—		—	585	73.76%	—	—	—	—	No
				PET	—						585	85.15%
				—	MRI+PET						585+585	92.34%
[61] (2020)	2D CNN	ROI based	Spatially segmented and normalized, skull stripping, coregistration	—	MRI+DTI	ADNI	115	106		185	406	93.50%	—	79.6%	—	—	No
[62] (2021)	3DCNN	Voxel based	Registration and segmentation	MRI	—	ADNI	146	146		256	548	—	89.3%	87.5%	—	No	No
[63] (2021)	DCNN, VGG-16, VGG-19	Slice based	Flipping, random zooming	MRI	—	OASIS	—	—		—	416	—	—	—	71%	Yes	Yes
[64] (2021)	Deep transfer ensemble (DTE)	Slice based	FWHM, segmentation, registration	MRI	—	ADNI	187	398		228	813	99%	98.7%	—	—	Yes	Yes
[65] (2021)	VGG; DenseNet; ResNet; EfficientNet	Slice based	—	MRI	—	OASIS	—	—		—	416	—	—	—	72%; 92%; 93%; 96%	Yes	Yes
[66] (2021)	AlexNet	Slice based	—	MRI	—	OASIS	—	—		—	664	—	—	—	96%	Yes	Yes
[67] (2021)	VGG-16	Slice based	3D to 2D conversion	fMRI	—	ADNI	18	—		36	54	99.9	—	—	—	Yes	Yes