Deep Grading based on Collective Artificial

Intelligence for AD Diagnosis and Prognosis

Huy-Dung Nguyen1 , Michaël Clément1 , Boris Mansencal1 , and Pierrick Coupé1

1
Univ. Bordeaux, CNRS, Bordeaux INP, LaBRI, UMR 5800, 33400 Talence, France

Abstract. Accurate diagnosis and prognosis of Alzheimer’s disease are

crucial to develop new therapies and reduce the associated costs. Re-
cently, with the advances of convolutional neural networks, methods have
been proposed to automate these two tasks using structural MRI. How-
ever, these methods often suffer from lack of interpretability, general-
ization, and can be limited in terms of performance. In this paper, we
propose a novel deep framework designed to overcome these limitations.
Our framework consists of two stages. In the first stage, we propose
a deep grading model to extract meaningful features. To enhance the
robustness of these features against domain shift, we introduce an inno-
vative collective artificial intelligence strategy for training and evaluating
steps. In the second stage, we use a graph convolutional neural network
to better capture AD signatures. Our experiments based on 2074 sub-
jects show the competitive performance of our deep framework compared
to state-of-the-art methods on different datasets for both AD diagnosis
and prognosis.

Keywords: Deep Grading · Collective Artificial Intelligence · General-

ization · Alzheimer’s disease classification · Mild Cognitive Impairment

1 Introduction
The first cognitive symptoms of Alzheimer’s disease (AD) appear right after
the morphological changes caused by brain atrophy [10]. Those changes can be
identified with the help of structural magnetic resonance imaging (sMRI) [2].
Recently, with the advances of convolutional neural networks (CNN), methods
have been proposed for automatic AD diagnosis using sMRI. Despite encourag-
ing results, current deep learning methods suffer from several limitations. First,
deep models lack transparency in their decision-making process [31, 38]. There-
fore, this limits their use for computer-aided diagnosis tools in clinical practice.
Second, for medical applications, the generalization capacity of classification
models is essential. However, only a few works have proposed methods robust
to domain shift [13, 34]. Third, current CNN models proposed for AD diagno-
sis and prognosis still perform poorly [35]. Indeed, when properly validated on
external datasets, current CNN-based methods perform worse than traditional
approaches (i.e., standard linear SVM).
In this paper, to address these three major limitations, we propose a novel
interpretable, generalizable and accurate deep framework. An overview of our
Fig. 1. Overview of our processing pipeline. The MRI image, its segmentation and the
deep grading map illustrated are from an AD subject.

proposed pipeline is shown in Figure 1. First, we propose a novel Deep Grading

(DG) biomarker to improve the interpretability of deep model outputs. Inspired
by the patch-based grading frameworks [4,12,32], this new biomarker provides a
grading map with a score between −1 and 1 at each voxel related to the alteration
severity. This interpretable biomarker may help clinicians in their decision and to
improve our knowledge on AD progression over the brain. Second, we propose an
innovative collective artificial intelligence strategy to improve the generalization
across domains and to unseen tasks. As recently shown for segmentation [6, 18],
the use of a large number of networks capable of communicating offers a better
capacity for generalization. Based on a large number of CNNs (i.e., 125 U-Nets),
we propose a framework using collective artificial intelligence efficient on different
datasets and able to provide accurate prognosis while trained for diagnosis task.
Finally, we propose to use a graph-based modeling to better capture AD signa-
ture using both inter-subject similarity and intrasubject variability. As shown
in [12], such strategy improves performance in AD diagnosis and prognosis.
In this paper, our main contributions are threefold:
– A novel deep grading biomarker providing interpretable grading maps.
– An innovative collective artificial intelligence strategy robust to unseen datasets
and unknown tasks.
– A new graph convolutional network (GCN) model for classification offering
state-of-the-art performance for both AD diagnosis and prognosis.

2 Materials and method

2.1 Datasets
The data used in this study, consisting of 2074 subjects, were obtained from mul-
tiple cohorts: the Alzheimer’s Disease Neuroimaging Initiative (ADNI) [16], the
Table 1. Number of participants used in our study. Data used for training is in bold.

Dataset CN AD sMCI pMCI

ADNI1 170 170 129 171
ADNI2 149 149 - -
AIBL 233 47 12 20
OASIS3 658 97 - -
MIRIAD 23 46 - -

Open Access Series of Imaging Studies (OASIS) [21], the Australian Imaging,
Biomarkers and Lifestyle (AIBL) [7], the Minimal Interval Resonance Imaging
in Alzheimer’s Disease (MIRIAD) [27]. We used the baseline T1-weighted MRI
available in each of these studies. Each dataset contains AD patients and cog-
nitively normal (CN) subjects. For ADNI1 and AIBL, it also includes mild cog-
nitive impairment (MCI), the early stage of AD composed of abnormal memory
dysfunctions. Two groups of MCI are considered: progressive MCI (pMCI) and
stable MCI (sMCI). The definition of these two groups is the same as in [35].
Table 1 summarizes the number of participants for each dataset used in this
study. During experiments, AD and CN from ADNI1 are used as training set
and the other subjects as testing set.

2.2 Preprocessing

All the T1w MRI are preprocessed using the following steps: (1) denoising [29],
(2) inhomogeneity correction [33], (3) affine registration into MNI space (181 ×
217 × 181 voxels at 1mm × 1mm × 1mm) [1], (4) intensity standardization [28]
and (5) intracranial cavity (ICC) extraction [30]. After that preprocessing, we
use AssemblyNet [6] to segment 133 brain structures (see Figure 1). The list of
structures is the same as in [14]. In this study, brain structure segmentation is
used to determine the structure volume (i.e., normalized volume in % of ICC)
and aggregate information in the grading map (see Section 2.3 and Figure 1).

2.3 Deep Grading for disease visualization

In AD classification, most of deep learning models only use CNN as binary

classification tool. In this study, we propose to use CNN to produce 3D maps
indicating where specific anatomical patterns are present and the importance of
structural changes caused by AD.
To capture these anatomical alterations, we extend the idea of the patch-
based grading (PBG) framework [4, 12, 32]. The PBG framework provides a 3D
grading map with a score between −1 and 1 at each voxel related to the alteration
severity. Contrary to previous PBG methods based on non-local mean strategy,
here we propose a novel DG framework based on 3D U-Nets.
Concretely, each U-Net (similar to [6]) takes a 3D sMRI patch (e.g., 32 ×
48 × 32) and outputs a grading map with values in range [−1, 1] for each voxel.
Voxels with a higher value are considered closer to AD, while voxels with a lower
value are considered closer to CN. For the ground-truth used during training,
we assign the value 1 (resp. −1) to all voxels inside a patch extracted from an
AD patient (resp. CN subject). All voxels outside of ICC are set to 0.
Once trained, the deep models are used to grade patches. These local outputs
are gathered to reconstruct the final grading map (see Section 2.4). Using the
structure segmentation, we represent each brain structure grading by its average
grading score (see Figure 1). This anatomically driven aggregation allows better
and meaningful visualization of the disease progression. In this way, during the
classification step (see Section 2.5), each subject is encoded by an n-dimensional
vector where n is the number of brain structures.

2.4 Collective AI for grading

As recently shown in [3, 35], current AD classification techniques suffer from a

lack of generalization. In this work, we propose an innovative collective artificial
intelligence strategy to improve the generalization across domains and to unseen
tasks. As recently shown for segmentation [6, 18], the use of a large number of
compact networks capable of communicating offers a better capacity for gener-
alization. There are many advantages to using the collective AI strategy. First,
it addresses the problem of GPU memory in 3D since each model processes only
a sub-volume of the image. The use of a large number of compact networks is
equivalent to a big neural network with more filters. Second, the voting system
based on a large number of specialized and diversified models helps the final
grading decision to be more robust against domain shift and different tasks.
Concretely, a preprocessed sMRI is decomposed into k × k × k overlapping
patches of the same size (e.g., 32 × 48 × 32). During training, for each patch
localization in the MNI space, a specialized model is trained. Therefore, in our
case (k = 5), we trained m = k × k × k = 125 U-Nets to cover the whole image
(see Fig. 1). Moreover, each U-Net is initialized using transfer learning from
its nearest neighbor U-Nets in the MNI space, except the first one trained from
scratch as proposed in [6]. As adjacent patches have some common patterns, this
communication allows grading models to share useful knowledge between them.
For each patch, 80% of the training dataset (i.e., ADNI1) is used for training
and the remaining 20% for validation. The accuracy obtained on validation set
is used to reconstruct the final grading map using a weighted average as follows:

P
xi ∈Pj αj ∗ gij
Gi = P (1)
xi ∈Pj αj

where Gi is the grading score of the voxel xi in the final grading map, gij is
the grading score of the voxel xi in the local grading patch Pj , and αj is the
validation accuracy of the patch j. This weighted vote enables to give more
weight to the decision of accurate models during the reconstruction.
2.5 Graph convolutional neural network for classification

The DG feature provides an inter-subject similarity biomarker which is helpful

to detect AD signature. However, the structural alterations leading to cognitive
decline could be different between subjects. Indeed, following the idea of [12],
we model the intra-subject variabilities by a graph representation to capture
the relationships between several regions related to the disease. We define an
undirected graph G = (N, E), where N = {n1 , . . . , ns } is the set of nodes for
the s brain structures and E = s × s is the matrix of edge connections. In our
approach, all nodes are connected with each other in a complete graph, where
nodes embed brain features (e.g., our proposed DG feature) and potentially other
types of external features.
Indeed, besides the grading map, the volume of structures obtained from the
segmentation could be helpful to distinguish AD patients from CN [12, 32]. It is
due to the evidence that AD leads to structure atrophy. Age is also an important
factor as, within sMRI, patterns in the brain of young AD patients could be
similar to elder CN. Indeed, the combination of those features is expected to
improve our classification performance. In our method, each node represents
a brain structure and embeds a feature vector (DG, V, A) where V and A are
respectively the volume of structures and subject’s age. Finally, we use the graph
convolutional neural network (GCN) [20] as the way to pass messages between
nodes and to perform final classification.

2.6 Implementation details

First, we downsample the sMRI from 181 × 217 × 181 voxels (at 1mm) to
91 × 109 × 91 voxels to reduce the computational cost, then decompose them
into 5 × 5 × 5 overlapping patches of size 32 × 48 × 32 voxels equally spaced
along the three axis. For each patch, an U-Net is trained using mean absolute
error loss, Adam optimizer with a learning rate of 0.001. The training process is
stopped after 20 epochs without improvement in validation loss. We employed
several data augmentation and sampling strategies to alleviate the overfitting
issue during training. A small perturbation is first created in training samples
by randomly translating by t ∈ {−1, 0, 1} voxel in 3 dimensions of the image.
We then apply the mixup [37] data augmentation scheme that was shown to
improve the generalization capacity of CNN in image classification.
Once the DG feature is obtained, we represent each subject by a graph of
133 nodes. Each node represents a brain structure and embeds DG, volume and
age features. Our classifier is composed of 3 layers of GCN with 32 channels,
followed by a global mean average pooling layer and a fully connected layer
with an output size of 1. The model is trained using the binary cross-entropy
loss, Adam optimizer with a learning rate of 0.0003. The training process is
stopped after 20 epochs without improvement in validation loss. At inference
time, we randomly add noise X ∼ N (0, 0.01) to the node features and compute
the average of 3 predictions to get the global decision. Experiments have shown
that it helps our GCN to be more stable.
 For training and evaluating steps, we use a standard GPU (i.e., NVIDIA
TITAN X) with 12Gb of memory.

3 Experimental results

In this study, the grading models and classifiers are trained using ADNI1 dataset
within AD and CN subjects. Then, we assess their generalization capacity in
domain shift using AD, CN subjects from ADNI2, AIBL, OASIS, MIRIAD. The
generalization capacity in derived tasks is performed using pMCI, sMCI subjects
from ADNI1 (same domain) and AIBL (out of domain).
Influence of collective AI strategy. In this part, the DG feature is de-
noted as DGC (resp. DGI ) when obtained with the collective (resp. individual)
AI strategy. The individual AI strategy refers to the use of a single U-Net to
learn patterns from all patches of sMRI. We compare the efficiency of DGC and
DGI feature when using the same classifier (i.e., SVM or GCN) (see Table 2).
These experiments show that using DGC achieves better results in most con-
figurations. When using SVM classifier, we observe a gain of 3.6% (resp. 0.8%)
on average in AD/CN (resp. pMCI/sMCI) classification. The efficiency of GC
feature is even better with GCN classifier, where a gain of 4.0% (resp. 3.5%) is
observed.
Influence of GCN classifier. Besides the DG feature, the intra-subject
variabilities are also integrated into our graph representation. Hence, it should
be beneficial to use GCN to exploit all this information. In our experiments,
GCN outperforms SVM in all the tests using either DGI or DGC feature (see
Table 2). Concretely, using DGI feature, we observe a gain of 5.0% (resp. 7.6%)
on average for AD/CN (resp. pMCI/sMCI) classification. These improvements
are 5.4% and 10.6% when using DGC feature.
Influence of using additional non-image features. Moreover, we ana-
lyze the model performance using DGC with the structural volume V and age
A as additional node features in our graph representation. By using the com-
bined features, the performance on average in AD/CN and pMCI/sMCI is both
improved by 0.3% and 1.4% compared to DGC feature (see Table 2). In the rest
of this paper, these results are used to compare with current methods.
Comparison with state-of-the-art methods. Table 3 summarizes the
current performance of state-of-the-art methods proposed for AD diagnosis and
prognosis classification that have been validated on external datasets. In this
comparison we considered five categories of deep methods: patch-based strategy
based on a single model (Patch-based CNN [35]), patch-based strategy based on
multiple models (Landmark-based CNN [24], Hierarchical FCN [23]), ROI-based
strategy based on a single model focused on hippocampus (ROI-based CNN [35]),
subject-based considering the whole image based on a single model (subject-
based CNN [35], 3D Inception-ResNet-v2 [26], Efficient 3D [36] and AD2 A [11])
and a classical voxel-based model using a SVM (Voxel-based SVM [35]).
For AD diagnosis (i.e., AD/CN), all the methods show good balanced ac-
curacy, although some of them failed to generalize on OASIS. In this scenario
Table 2. Validation of the collective AI strategy, GCN classifier, the combination
of DG feature with other image and non-image features using GCN classifier. Red:
best result, Blue: second best result. The balanced accuracy (BACC) is used to assess
the model performance. The results are the average accuracy of 10 repetitions and
presented in percentage. All the methods are trained on the AD/CN subjects of the
ADNI1 dataset.

AD/CN pMCI/sMCI Average

Classifier Features
ADNI2 OASIS MIRIAD AIBL ADNI1 AIBL AD/CN p/sMCI
SVM DGI 83 83 88 79 65 66 83.3 65.5
SVM DGC 83 84 91 87 68 64 86.3 66.0
GCN DGI 84 88 96 82 68 73 87.5 70.5
GCN DGC 87 89 100 88 70 76 91.0 73.0
GCN DGC , V, A 87 88 98 92 74 74 91.3 74.0

Table 3. Comparison of our method with current methods in AD diagnosis and prog-
nosis. Red: best result, Blue: second best result. The balanced accuracy (BACC) is used
to assess the model performance. All the methods are trained on the AD/CN subject
of the ADNI1 dataset (except [23] that is fined-tuned on MCI subjects for sMCI/pMCI
task).

AD/CN pMCI/sMCI
Methods
ADNI2 OASIS MIRIAD AIBL ADNI1 AIBL
Landmark-based CNN [24] 91 - 92 - - -
Hierachical FCN [23] 89 - - - 69 -
Patch-based CNN [35] - 64 - 81 70 64
ROI-based CNN [35] - 69 - 84 70 60
Subject-based CNN [35] - 67 - 83 69 52
Voxel-based SVM [35] - 70 - 88 75 62
AD2 A [11] 88 - - 88 - -
Efficient 3D [36] - 92 96 91 70 65
3D Inception-ResNet-v2 [26] - 85 - 91 42 -
Our method 87 88 98 92 74 74

(unseen datasets), our method obtained high accuracy for all the datasets. This
confirms the generalization capacity of our approach against domain shift.
For AD prognosis (i.e., pMCI/sMCI), we observe a significant drop for all
the methods. This drop is expected since pMCI/sMCI classification is more
challenging and since models are trained on a different task (i.e., AD/CN). For
this task, our method is generally robust, especially on AIBL. Moreover, our
approach is the only deep learning method that performs competitively with the
SVM model [35] on ADNI1, while significantly better on AIBL. In this scenario
(unknown task), our method obtains the highest accuracy on average. These
results highlight the potential performance of our method on unseen tasks.
Interpretation of collective deep grading. To highlight the interpretabil-
ity capabilities offered by our DG feature, we first compute the average DG map
for each group: AD, pMCI, sMCI and CN (see Figure 2). First, we can note that
the average grading maps increase between each stage of the disease. Second, we
estimated the top 10 structures with highest absolute value of grading score over
 Fig. 2. Average grading map per group of subjects.

all the testing subjects. The found structures are known to be specifically and
early impacted by AD. These structures are: bilateral hippocampus [9], left amyg-
dala and left inferior lateral ventricle [5], left parahippocampal gyrus [19], left
posterior insula [8], left thalamus proper [17], left transverse temporal gyrus [25],
left ventral diencephalon [22]. While other attention-based deep methods failed
to find structures related to AD [3], our DG framework shows high correlation
with current physiopathological knowledge on AD [15].

4 Conclusion

In this paper, we addressed three major limitations of CNN-based methods by

introducing a novel interpretable, generalizable and accurate deep grading frame-
work. First, deep grading offers a meaningful visualization of the disease progres-
sion. Second, we proposed a collective artificial intelligence strategy to improve
the generalization of our DG strategy. Experimental results showed a gain for
both SVM and GCN in all tasks using this strategy. Finally, we proposed to use
a graph-based modeling to better capture AD signature using both inter-subject
similarity and intra-subject variability. Based on that, our DG method showed
state-of-the-art performance in both AD diagnosis and prognosis.

Acknowledgments This work benefited from the support of the project Deepvol-
Brain of the French National Research Agency (ANR-18-CE45-0013). This study was
achieved within the context of the Laboratory of Excellence TRAIL ANR-10-LABX-57
for the BigDataBrain project. Moreover, we thank the Investments for the future Pro-
gram IdEx Bordeaux (ANR-10-IDEX-03-02), the French Ministry of Education and
Research, and the CNRS for DeepMultiBrain project.
References
1. Avants, B.B., et al.: A reproducible evaluation of ANTs similarity metric perfor-
mance in brain image registration. NeuroImage 54(3), 2033–2044 (2011)
2. Bron, E., et al.: Standardized evaluation of algorithms for computer-aided diagnosis
of dementia based on structural MRI: The CADDementia challenge. NeuroImage
111, 562–579 (2015)
3. Bron, E., et al.: Cross-cohort generalizability of deep and conventional machine
learning for MRI-based diagnosis and prediction of Alzheimer’s disease. NeuroIm-
age: Clinical 31, 102712 (2021)
4. Coupé, P., et al.: Scoring by nonlocal image patch estimator for early detection of
Alzheimer’s disease. NeuroImage: Clinical 1(1), 141–152 (2012)
5. Coupé, P., et al.: Lifespan changes of the human brain in Alzheimer’s disease.
Nature Scientific Reports 9(3998) (2019)
6. Coupé, P., et al.: AssemblyNet: A large ensemble of CNNs for 3D whole brain MRI
segmentation. NeuroImage 219, 117026 (2020)
7. Ellis, K., et al.: The Australian Imaging, Biomarkers and Lifestyle (AIBL) study of
aging: Methodology and baseline characteristics of 1112 individuals recruited for
a longitudinal study of Alzheimer’s disease. International psychogeriatrics / IPA
21, 672–87 (2009)
8. Foundas, A., et al.: Atrophy of the hippocampus, parietal cortex, and insula in
alzheimer’s disease: a volumetric magnetic resonance imaging study. Neuropsychi-
atry, neuropsychology, and behavioral neurology 10(2), 81—89 (1997)
9. Frisoni, G., et al.: The clinical use of structural MRI in Alzheimer’s disease. Nature
reviews. Neurology 6, 67–77 (2010)
10. Gordon, B., et al.: Spatial patterns of neuroimaging biomarker change in individ-
uals from families with autosomal dominant Alzheimer’s disease: A longitudinal
study. The Lancet Neurology 17 (2018)
11. Guan, H., et al.: Attention-guided deep domain adaptation for brain dementia iden-
tification with multi-site neuroimaging data. In: MICCAI Workshop on Domain
Adaptation and Representation Transfer (2020)
12. Hett, K., et al.: Graph of brain structures grading for early detection of Alzheimer’s
disease. In: Medical Image Computing and Computer-Assisted Intervention (MIC-
CAI) (2018)
13. Hosseini-Asl, E., et al.: Alzheimer’s disease diagnostics by adaptation of 3D convo-
lutional network. In: IEEE International Conference on Image Processing (ICIP)
(2016)
14. Huo, Y., et al.: 3D whole brain segmentation using spatially localized atlas network
tiles. NeuroImage 194, 105–119 (2019)
15. Jack, C., et al.: Hypothetical model of dynamic biomarkers of the Alzheimer’s
pathological cascade. The Lancet Neurology 27, 685–691 (2010)
16. Jack Jr., C.R., et al.: The Alzheimer’s disease neuroimaging initiative (ADNI):
MRI methods. Journal of Magnetic Resonance Imaging 27(4), 685–691 (2008)
17. de Jong, L.W., et al.: Strongly reduced volumes of putamen and thalamus in
Alzheimer’s disease: An MRI study. Brain 131(12), 3277–3285 (2008)
18. Kamraoui, R.A., et al.: Towards broader generalization of deep learning methods
for multiple sclerosis lesion segmentation. arXiv 2012.07950 (2020)
19. Kesslak, J.P., et al.: Quantification of magnetic resonance scans for hippocam-
pal and parahippocampal atrophy in alzheimer’s disease. Neurology 41(1), 51–51
(1991)
20. Kipf, T., et al.: Semi-supervised classification with graph convolutional networks.
In: International Conference on Learning Representations (ICLR) (2017)
21. LaMontagne, P.J., et al.: OASIS-3: Longitudinal Neuroimaging, Clinical, and Cog-
nitive Dataset for Normal Aging and Alzheimer Disease. medRxiv (2019)
22. Lebedeva, A.K., et al.: MRI-based classification models in prediction of mild cog-
nitive impairment and dementia in late-Life depression. Frontiers in Aging Neuro-
science 9, 13 (2017)
23. Lian, C., et al.: Hierarchical fully convolutional network for joint atrophy localiza-
tion and Alzheimer’s disease diagnosis using structural MRI. IEEE Transactions
on Pattern Analysis and Machine Intelligence 42(4), 880–893 (2020)
24. Liu, M., et al.: Landmark-based Deep Multi-Instance Learning for Brain Disease
Diagnosis. Medical Image Analysis 43 (2017)
25. Liu, Y., et al.: Education increases reserve against Alzheimer’s disease-Evidence
from structural MRI analysis. Neuroradiology 54, 929–38 (2012)
26. Lu, B., et al.: A practical Alzheimer disease classifier via brain imaging-based deep
learning on 85,721 samples. bioRxiv (2021)
27. Malone, I.B., et al.: MIRIAD—Public release of a multiple time point Alzheimer’s
MR imaging dataset. NeuroImage 70, 33–36 (2013)
28. Manjón, J.V., et al.: Robust MRI brain tissue parameter estimation by multistage
outlier rejection. Magnetic Resonance in Medicine 59(4), 866–873 (2008)
29. Manjón, J.V., et al.: Adaptive non-local means denoising of MR images with spa-
tially varying noise levels. Journal of Magnetic Resonance Imaging 31(1), 192–203
(2010)
30. Manjón, J.V., et al.: NICE: Non-local Intracranial Cavity Extraction. International
Journal of Biomedical Imaging (2014)
31. Nigri, E., et al.: Explainable deep CNNs for MRI-based diagnosis of Alzheimer’s
disease. In: International Joint Conference on Neural Networks (IJCNN) (2020)
32. Tong, T., et al.: A novel grading biomarker for the prediction of conversion from
mild cognitive impairment to alzheimer’s disease. IEEE Transactions on Biomedical
Engineering (2016)
33. Tustison, N.J., et al.: N4ITK: Improved N3 bias correction. IEEE Transactions on
Medical Imaging 29(6), 1310–1320 (2010)
34. Wachinger, C., et al.: Domain adaptation for Alzheimer’s disease diagnostics. Neu-
roImage 139 (2016)
35. Wen, J., et al.: Convolutional neural networks for classification of Alzheimer’s
disease: Overview and reproducible evaluation. Medical Image Analysis 63, 101694
(2020)
36. Yee, E., et al.: Construction of MRI-based Alzheimer’s disease score based on effi-
cient 3D convolutional neural network: Comprehensive validation on 7,902 images
from a multi-center dataset. Journal of Alzheimer’s Disease 79, 1–12 (2020)
37. Zhang, H., et al.: mixup: Beyond Empirical Risk Minimization. In: International
Conference on Learning Representations (ICLR) (2018)
38. Zhang, X., et al.: An explainable 3D residual self-attention deep neural network
for joint atrophy localization and Alzheimer’s Disease diagnosis using structural
MRI. IEEE Journal of Biomedical and Health Informatics (2021)