See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/336641380

Alzheimer's Disease Diagnosis Using Enhanced Inception Network Based on

Brain Magnetic Resonance Image

Conference Paper · October 2019

DOI: 10.1109/BIBM47256.2019.8983046

CITATIONS READS

0 108

6 authors, including:

Zhenyu Cui Jiaxu Leng

Chinese Academy of Sciences Chinese Academy of Sciences
7 PUBLICATIONS   4 CITATIONS    17 PUBLICATIONS   31 CITATIONS   

SEE PROFILE SEE PROFILE

Tianlin Zhang Pei Quan

Chinese Academy of Sciences Chinese Academy of Sciences
12 PUBLICATIONS   7 CITATIONS    11 PUBLICATIONS   7 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

object detection View project

UCAS_HPDM View project

All content following this page was uploaded by Jiaxu Leng on 18 October 2019.

The user has requested enhancement of the downloaded file.

 Alzheimer’s Disease Diagnosis Using Enhanced
Inception Network Based on Brain Magnetic
Resonance Image
Zhenyu Cui∗† , Zhiao Gao‡ , Jiaxu Leng∗† , Tianlin Zhang∗† , Pei Quan∗† , Wei Zhao∗
∗ School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing, 100190 China
† Key Lab of Big Data Mining and Knowledge Management, Chinese Academy of Sciences, Beijing, 100190 China
‡ School of Computer and Information Security, Guilin University of Electronic Technology, 541004 China

Corresponding author: Jiaxu Leng

Abstract—An estimated 24 million people worldwide have sessment) [8], 2)diagnosing based on the structural Magnetic
dementia, the majority of whom are thought to have Alzheimer’s Resonance(MR) images [9], 3)cerebrospinal fluid biomarker
disease(AD). Nowadays, Alzheimer’s disease represents a signifi- testing [10].
cant public health concern and has been identified as a research
priority. Most unfortunately, there is little chance of a cure for However, the diagnosis of Alzheimer’s disease is a very
Alzheimer’s disease, and the disease is difficult to detect before cumbersome process regardless of adopting any of the above
the dominant characteristics such as memory loss are manifested. methods. We believe that if there is a faster and more accurate
Therefore, the diagnosis of Alzheimer’s disease has become an diagnosing method, it will greatly improve the efficiency and
urgent problem today. Studies have shown that mild cognitive accuracy of the diagnosis and reduce the pressure on doc-
impairment(MCI) is a state between Alzheimer’s disease and nor-
mal, and the chance of it turning into Alzheimer’s disease is high. tors. Although the neuroimaging-based Alzheimer’s disease
Therefore, if machines can automatically learn the characteristics diagnosis method proposed in this paper cannot be used as
of three kinds of human brain magnetic resonance(MR) images the main diagnosis method, which can be only used as an
through deep learning, and help doctors to diagnose patients with auxiliary method, it is a faster and efficient auxiliary diagnosis
mild cognitive impairment or Alzheimer’s disease accurately, it method. By performing a thin-layer scan of the subjects’ heads
will be beneficial for the early diagnosis of Alzheimer’s disease.
In this paper, we improve the Inception(V3) neural network and and examining the coronal position, it is possible to show a
further test the effectiveness of the enhanced network based on marked atrophy of the cerebral cortex, especially the status
the international Alzheimer’s disease data set, which consists of of the hippocampus and the medial temporal lobe, which
brain magnetic resonance images. The results show that the total can be a strong support for the diagnosis of Alzheimer’s
accuracy of our approach can reach 85.7%. disease [1]. The current MR images cannot be the main tool
Keywords—Alzheimer’s disease diagnosis, magnetic resonance
image, neural network for diagnosis because people could hardly analyze the early
brain lesions of Alzheimer’s disease patients from them [11].
Nevertheless, if we can use the ability of deep learning to
I. I NTRODUCTION
classify images effectively and design a deep learning-based
Modern medicine has developed rapidly with the continu- diagnosis of Alzheimer’s disease, which is to analyze the MR
ous advancement of human civilization. Some diseases that images, it will undoubtedly contribute to the early diagnosis
were previously regarded as incurable nightmares have been of the disease and timely follow-up treatment.
overcome by scientists step by step. Alzheimer’s disease is the Image classification technology can get different feedbacks
most common cause of dementia and has not been effectively according to different features of pictures with different cat-
cured so far [1]. However, it can be certain that as long as egories. It is an image processing method that can classify
given a little more time, the cure of Alzheimer’s disease will different categories of pictures. This quantitative analysis
be just around the corner. Nowadays, the society shows a method extracts the features of the pixels in images before
high incidence trend in Alzheimer’s disease [2]. If we can performing tasks such as object recognition, detection, and
diagnose the early stage of Alzheimer’s disease, mild cognitive segmentation based on the features, which is also generally
impairment, and treat it with reasonable medical to delay the considered to replace the human visual judgment.
disease, modern medicine will be given more time to save Although humans can recognize more types of image recog-
these suffering old people [3]. In fact, when certain diseases nition work based on various knowledge, and fully consider
cannot be cured currently, not only Alzheimer’s disease [4] [5], the contextual information of the image, as of 2017, the
what can be done is to delay the time of onset of such diseases classification accuracy of image classification algorithms in
or to diagnose early [6] [7]. At present, matured diagnosis fixed tasks has exceeded that of human’s [12], which gives
methods of Alzheimer’s disease are as follows: 1)scoring us confidence in migrating the recognition technology to
test based on GPCOG (General Practitioner Cognitive As- diagnosing Alzheimer’s disease.
 Convolutional neural networks (CNN), which have been
proven to be effective in tasks such as object recognition [13], Image preprocess Recognition&Diagnosis
object detection [14], and object segmentation [15], are gradu-
MRI Images
ally becoming the mainstream method in dealing with complex
image processing tasks. In 2012, Krizhevsky proposed a new MRI Images for
Test
Testing
convolutional neural network [16] that made subversive ad-
Histogram
vances in the ImageNet target recognition competition that laid Equalization Enhanced
Inception Diagnosis
Result
the foundation for deep learning in image processing. In recent Neural
Network
years, with the gradual improvement of parallel computing
Fluid/Non-Fluid MRI Images for
Train
power, deeper convolutional neural networks have created Segmentation Training with Label
new state-of-the-art in different tasks in the field of computer
vision continuously. However, GoogleNet [17] has proposed
a new strategy: to expand the width of neural networks. Fig. 1: The architecture of our Alzheimer’s disease diagnosis.
GoogleNet’s goal is to design an excellent local topology,
which obtains multiple feature vectors by performing multiple
convolution operations in the same layer parallel, and finally output is diagnosis results containing the type of disease, in-
concatenate these feature vectors to obtain a very deep feature cluding: no Alzheimer’s disease, suffering from mild cognitive
map. Such stacking feature maps can better reflect more impairment(MCI) and suffering from Alzheimer’s disease. The
features of the image in a single convolutional block, which is first part of our approach is to pre-process the MR images
called Inception. Inspired by the successful application of the where we creatively proposed a method combining the OTSU
network structure [17], we use the Inception(V3) network as [18] and the histogram equalization [19]. The second part uses
the main body of our network, and enhance the network for the merged characteristic images after processing the first part
better adaption of the diagnosis of Alzheimer’s disease. This for model training and forward prediction. In this part, we
paper reveals our method for the diagnosis of Alzheimer’s optimize the traditional Inception(V3) network properly and
disease based on deep learning. The theme of the network is finally improve the diagnosis of Alzheimer’s disease.
the Inception(V3) [17] neural network, which reads the MR
images through files and then feeds the processed images to B. Image Pre-Processing
the deep learning model. The neural network will use the Currently, MR images are widely used in the pathological
classification result as the diagnosis result of Alzheimer’s analysis [20]. We have found that different MR images tend
disease. The main contributions of this paper are as follows: to have different shades. We believe that this is closely related
1) The application of deep neural network to the diagnosis to the amount of radiation when imaging and the intensity of
of Alzheimer’s disease. the backlight when they are displayed. As a result, the image
2) The pre-processing method of MR image for pixel values are unevenly distributed, which has a negative
Alzheimer’s disease. impact on the training of neural networks [21]. Therefore,
3) The enhancement of Inception(V3) network to improve we first use the histogram equalization to generate a contrast-
the accuracy of the whole model prediction and diagno- adjusted image C(x, y) for the original MR image G(x, y).
sis. Histogram equalization is an image contrast adjusting method
The rest of this paper are organized as follows. Section II [19], and the adjusted image have a wider range of gray value
discusses the details of our approach. Section III presents and distribution than before.
analyzes experimental results. Section IV conclude the whole (
G(x, y) O(x, y) = v
work of this paper. Gv (x, y) = (1)
0 Else
II. O UR APPROACH (
C(x, y) O(x, y) = v
In this section, we introduce our approach of Alzheimer’s Cv (x, y) = (2)
disease diagnosis based on the enhanced Inception(V3) net- 0 Else
work. Subsequently, each module in the method is discussed In addition, the MR image is unique in that the flowing
in detail in the following subsections. liquid does not produce signals, called “The Flowing Effect”
[22]. This makes the signals brought by structures such as
A. Overview of Our Framework bones and blood vessels are stronger. Meanwhile, the blood
The architecture of our proposed method for the diagnosis and the flowing liquid look more black on MR images which
of Alzheimer’s disease using MR image as shown in Fig.1. represents no signal. Based on this knowledge, bones and
Our diagnosis framework consists of two parts. One is soft tissues can be easily separated by humans. However, the
the pre-processing part of MR images, and the other is the contrast of this relative black and white is not strong enough
recognition part based on the enhanced Inception(V3) neural for machines. We believe that if these two parts can be used in-
network. The input to our method is MR images, while the dependently in different input channels of the neural network,
 OUT OUT OUT

n×1

1×n n×1 3×3 1×3 3×1

1×1 n×1 1×n 1×1 3×3 3×3 1×1 1×3 3×1 3×3

Pool 1×n 1×1 1×1 Pool 1×1 1×1 1×1 Pool 1×1 1×1 1×1

1×1

BASE BASE BASE

(a) (b) (c)

Fig. 2: Three inception block structures in Inception(V3) network.

it can minimize the interference of different information of TABLE I: Network structure of Inception(V3) network
issues on neural network learning. Therefore, we draw on the type patch size/stride input size
approach of OTSU [18] to identify organizations with different or remarks
conv 3*3/2 299*299*3
mobile states with foreground which value is 0 and background conv 3*3/1 149*149*32
which value is 1 to form the graph O(x, y). For different conv padded 3*3/1 147*147*32
annotations in O(x, y), we divide the corresponding pixels in pool 3*3/2 147*147*64
G(x, y) and C(x, y) into two parts, namely G0 (x, y), G1 (x, y) conv 3*3/1 73*73*64
conv 3*3/2 71*71*80
and C0 (x, y), C1 (x, y), as shown in Equation 1 and Equation conv 3*3/1 35*35*192
2. 3*Inception As in Fig.2(a) 35*35*288
5*Inception As in Fig.2(b) 17*17*768
2*Inception As in Fig.2(c) 8*8*1280
OU T = ST ACK(G, C, G0 , G1 , C0 , C1 ) (3)
pool 8*8 8*8*2048
Finally, we stack G(x, y), C(x, y), G0 (x, y), G1 (x, y), linear logits 1*1*2048
softmax classifier 1*1*1000
C0 (x, y) and C1 (x, y), and get the first part of the output of
the image, as shown in Equation 3. The pre-processed MR
image is changed from the original 3(RGB) channels to a
Combination”, combines multi-dimensional attentions just as
series feature maps with 3 ∗ 6 = 18 channels. We take these
doctors extract important information when diagnosing. Atten-
feature maps as the input of the second part of the enhanced
tion mechanisms have been widely used in pattern recognition
Inception(V3) neural network for training and testing.
in recent years [24], and its effectiveness has been proven
C. Inception(v3) Network repeatedly. The “Multi-Attention Combination” is combined
Unlike traditional deep neural network, the Inception(V3) with the characteristics of Alzheimer’s disease and is therefore
network is known for its parallel stacked convolutional layers. more targeted.
Table I is the parameters of the original network structure of First of all, the traditional diagnosis of Alzheimer’s dis-
Inception(V3) network. The neural network layers in the table ease is based on the experience of a doctor. When doctors
are connected in series from top to bottom. diagnosing based on MR images, they tend to pay more
Among Table I, three different inception block structures attention to the places that are decisive for diagnoses, such as
are shown in Fig.2. Inception(V3) network is the second hippocampus, frontal lobe and temporal lobe [20]. Therefore,
improvement of Inception neural network. Compared to the in order to guide the inception modules to learn these key areas
previous works [23] [21], Inception(V3) replaces the original independently during the network training process, we perform
n ∗ n convolution kernels by using 1 ∗ n and n ∗ 1 convolution two additional convolution layers (whose kernels are 3*3) on
kernels sequentially, which not only effectively improves the each input of the inception modules (the “BASE” layers in
network identification accuracy, but also reduces the number Fig.2), generating the corresponding attention heat maps which
of parameters and the time cost on network training. have the same size as the “OUT” layers in Fig.2 and single
channel. We define that all values have a range of (0, 1) in this
D. Enhanced Inception(V3) Network heat map, which is activated by using sigmoid [25] activation
Our improvement is based on the inception modules above. function and the closer the value on the heat map is to 1, the
The improved method we proposed, called “Multi-Attention more this location is concerned by the network. Finally, we
 multiply these coefficients by the corresponding channels, so
OUT that channels that are not worthy of attention are “ignored” to
achieve the purpose of compressing the number of channels.
Finally, our improved Inception module is shown in Fig.4.
OUT’
TABLE II: Network structure of enhanced Inception(V3)
network
n×1
type patch size/stride input size
remarks
sigmoid 1×n n×1 conv 3*3/2 192*160*18
conv 3*3/1 95*79*32
conv/padding 3*3/1 93*77*32
3×3 1×1 n×1 1×n
conv 1*1/1 93*77*64
conv 3*3/1 93*77*80
3×3 Pool 1×n 1×1 1×1 max pool 3*3/2 91*75*192
3*Inception As in Fig.4(a) 45*37*192
5*Inception As in Fig.4(b) 45*37*288
1×1 2*Inception As in Fig.4(c) 22*18*768
avg pool 10*8 10*8*2048
linear logits 1*1*2048
BASE softmax classifier 1*1*3

(a) Finally, our network structure is shown in the Table II. The
approach is not only applicable to Inception(V3) networks,
Fig. 3: The improved Inception(V3) blocks(a). but also suitable for any deep convolutional neural network
architecture. It is worth mentioning that the approach does not
increase the depth of neural network and is easy to deploy.
multiply the attention heat map by the corresponding position Considering that the identification type of this task is quite
of each channel of the “OUT” layer to obtain the feature maps different from the original network, we have modified the
that have been focused on. partial convolution and pooling operations of the Inception(v3)
In other words, the more focused on the pixel, the more network appropriately.
completely retained the feature value, and vice versa. Thus, in
the process of gradient descent(network training) [26], when
the classification is correct, the weight of the feature will be III. E XPERIMENT & E VALUATION
raised, and vice versa. Eventually, the network reduces the
attention values in the irrelevant part and learn the features A. Data Set
that are useful for classification. Fig.3 is the improved version
of Inception(V3) module(a) in Fig.2. To validate our approach, we experiment with an interna-
Secondly, the inception module is characterized by the tional MR image data set [27] provided by Alzheimers Disease
inclusion of convolution kernels of different scales in the Neuroimaging Initiative(ADNI) [28]. The data set is labeled as
same convolutional module [17]. The benefits are obvious: three categories: no Alzheimer’s disease, suffering from mild
Compared to a single convolution kernel, the inception module cognitive impairment and suffering from Alzheimer’s disease,
can extract multiple types of features from single-layer feature which includes a total of 662 three-dimensional brain MR im-
maps using a variety of convolution kernels, which expands the ages. Based on previous knowledge [20], the most significant
dimensions of each layer of feature maps without deepening difference between Alzheimer’s patients and healthy people is
the neural network. We believe that this module brings a new in the hippocampus, frontal lobe and some other areas. These
problem as well: “the feature map dimension disaster”. Feature regions are relatively concentrated in MR images. Therefore,
maps of thousands of channels appear in the final inception we have captured all MR images with a depth of 92 to 108,
layer Fig.2(c). However, these channels are not fully utilized, which contains all of the above areas. Then, we converted
just as doctors could hardly consider so many channels when these images into three-channel images, and we got a total
diagnosing Alzheimer’s disease. So, we introduce the “channel of 9952 two-dimensional MR images with RGB channels.
attention mechanism” into the Inception(V3) network to let the Fig.5 is a sample of grayscale images and the corresponding
network learn the importance of the channel in the process converted images.
of network training. Here, an important channel refers to the We randomly divide the data set into training set, validation
channel that can be used as a vital feature for classification. set and test set in a ratio of approximately 8:1:1. The grayscale
We generate a one-dimensional vector, whose values have a MR image size is 192*160 in the data set. We modified the
range of (0, 1), with the same length of output feature(“OUT” input of the original Inception(V3) network, which used to be
layer in Fig.2) channels through a fully connected layer. And 299*299, to overcome the effects of image distortion.
 OUT OUT OUT

...
1×n
OUT’’ OUT’’ OUT’’

...

...
1×n 1×n

sigmoid
sigmoid sigmoid

OUT’ fully connected OUT’ fully connected OUT’ fully connected

n×1

sigmoid 1×n n×1 sigmoid 3×3 sigmoid 1×3 3×1

3×3 1×1 n×1 1×n 3×3 1×1 3×3 3×3 3×3 1×1 1×3 3×1 3×3

3×3 Pool 1×n 1×1 1×1 3×3 Pool 1×1 1×1 1×1 3×3 Pool 1×1 1×1 1×1

1×1

BASE BASE BASE

(a) (b) (c)

Fig. 4: Three enhanced inception block structures we proposed.

TP
Sensitivity = (5)
TP + FN
To illustrate the effectiveness of our approach, we trained
the enhanced Inception(V3) network and tested our model on
the MR image data set [27]. The results are shown in Table
IV.
TABLE IV: The accuracy of our approach.
Categories Accuracy
Alzheimer’s Disease(AD) 1.000
Mild Cognitive Impairment(MCI) 0.895
Fig. 5: A sample of the data set [27]. Normal(NL) 0.770
TOTAL 0.857

B. Evaluation
TABLE V: The specificity and sensitivity of our approach.
The validation set contains 896 MR images, of which 176
Group Specificity Sensitivity
were labeled for suffering from Alzheimer’s disease, 304 for AD & NL 0.9375 1.000
mild cognitive impairment and 416 for normal. MCI & NL 0.8182 0.8947
We use accuracy, sensitivity, and specificity to evaluate our
approach. We evaluated the image which labeled for normal The results in the table indicates that our novel network has
and the image labeled for different types of disease separately. an accuracy of 85.7% on the verification set. Consequently, we
For each group of evaluation, the parameters of confusion calculate the specificity and sensitivity of the result for further
matrix are shown in Table III. Similarly, the calculation of the analysis as shown in Table V.
specificity as shown in Equation 4 and sensitivity as shown in Compared with the traditional method [29], our method is
Equation 5 rate can be derived from the confusion matrix. superior to the manual diagnosis method in both specificity
and sensitivity. However, this comparison is not completely
TABLE III: The confusion matrix for evaluation
sufficient, considering that the quantity and quality of the
Positive Label Negative Label information acquired by people in different eras are far apart.
Positive Prediction True Positive(TP) False Positive(FP) Therefore, we select some of the classic neural networks and
Negative prediction False Negative(FN) True Negative(TN)
verified them with the same data. The results are shown in
Table VI.
TN The experimental results show that our approach achieves
Specif icity = (4) 100% accuracy in the recognition of Alzheimer’s disease, and
FP + TN
 TABLE VI: The performance of different neural networks. [2] P. Schelterns, H. Feldman et al., “Treatment of
Neural Network AD & NL MCI & NL AD & NL & MCI alzheimer’s disease; current status and new perspectives,”
ResNet-101 [30] 0.9745 0.8555 0.8351 The Lancet Neurology, vol. 2, no. 9, pp. 539–547, 2003.
SVM [31] 0.9016 0.8026 0.7526
AlexNet [16] 0.9349 0.8304 0.8030
[3] P. Axelsson, J. M. Albandar, and T. E. Rams, “Prevention
VGG16 [32] 0.9583 0.8416 0.8125 and control of periodontal diseases in developing and
Inception(V3) [17] 0.9786 0.8605 0.8357 industrialized nations,” Periodontology, vol. 29, no. 1,
OUR’s 1.0000 0.8947 0.8571 pp. 235–246, 2010.
[4] P. Greenwald, C. K. Clifford, and J. A. Milner, “Diet and
cancer prevention,” European Journal of Cancer, vol. 37,
is superior to the current classical neural networks. The overall no. 8, pp. 948–965, 2001.
accuracy of 85.7% is also significantly higher than other [5] L. Bolondi, S. Sofia, S. Siringo, S. Gaiani, A. Casali,
methods, which fully proves the effectiveness of our novel G. Zironi, F. Piscaglia, L. Gramantieri, M. Zanetti, and
network. Meanwhile, we find an interesting phenomenon: no M. Sherman, “Surveillance programme of cirrhotic pa-
matter which method, the recognition accuracy of normal tients for early diagnosis and treatment of hepatocellular
people is lower than the average one, while the accuracy of carcinoma: a cost effectiveness analysis,” Gut, vol. 48,
Alzheimer’s disease is much higher than the average, which no. 2, pp. 251–259, 2001.
means a normal person may be misdiagnosed as a patient, but [6] D. J. Selkoe, “Preventing alzheimer’s disease,” Science,
the patient will hardly miss the diagnosis. From a pathological vol. 337, no. 6101, p. 1488, 2012.
point of view, if human brain tissues structural changed in [7] G. Serge, R. Barry, Z. Michael, R. C. Petersen, R. Karen,
MR images, there is a high probability of suffering from B. Karl, B. Sylvie, B. Henry, B. David, and C. Howard,
Alzheimer’s disease. However, brain tissues of normal people “Mild cognitive impairment,” Lancet, vol. 367, no. 9518,
are structurally diverse, and tissues that appear to have struc- p. 1262, 2006.
tural defects may not be functionally defective. We believe [8] H. Brodaty, D. Pond, N. M. Kemp, G. Luscombe,
that the various brain tissues of normal people is the culprit, L. Harding, K. Berman, and F. A. Huppert, “The gpcog:
and it might be difficult for the neural network to exclude a new screening test for dementia designed for general
brain tissue that looks like suffering from Alzheimer’s disease practice,” Journal of the American Geriatrics Society,
through MR images merely, which reveals that our method vol. 50, no. 3, pp. 530–534, 2002.
can only be used as a reference. The more comprehensive [9] NeuroImage, “Alzheimer’s disease diagnosis in individ-
diagnosis means the more types of diagnostic methods. ual subjects using structural mr images: Validation stud-
Taking a step back, we still insist that our approach is ies,” Neuroimage, vol. 39, no. 3, pp. 1186–1197, 2008.
effective. For diseases such as Alzheimer’s disease, missed [10] H. Vanderstichele, M. Bibl, S. Engelborghs, N. L. Bas-
detection is far more serious than false detection, and our tard, P. Lewczuk, J. L. Molinuevo, L. Parnetti, A. Perret-
method has a low rate of missed detection. In all, the approach Liaudet, L. M. Shaw, and C. Teunissen, “Standardization
proposed in this paper is a novel and effective method for the of preanalytical aspects of cerebrospinal fluid biomarker
diagnosis of Alzheimer’s disease. testing for alzheimer’s disease diagnosis: A consensus
paper from the alzheimer’s biomarkers standardization
IV. C ONCLUSION
initiative,” Alzheimers Dementia the Journal of the
This paper proposed an enhanced Inception(V3) neural Alzheimers Association, vol. 8, no. 1, pp. 65–73, 2012.
network for diagnosis of Alzheimer’s disease based on MR im- [11] J. W. Pettegrew, . Moossy, J., . Withers, G., . Mckeag, D.,
ages. We have proposed three brand new enhanced Inception and . Panchalingam, K., “31p nuclear magnetic resonance
blocks, which can be used to improve the recognition accuracy study of the brain in alzheimer’s disease,” Journal of
of Inception(V3) network and make diagnosis effectively. An Neuropathology Experimental Neurology, vol. 47, no. 3,
appropriate data pre-processing scheme of MR images has p. 235, 1988.
also been applied to the diagnosis of Alzheimer’s disease. [12] H. Jie, S. Li, S. Albanie, S. Gang, and E. Wu, “Squeeze-
Finally, after a series of comparisons and modifications, the and-excitation networks,” IEEE Transactions on Pattern
enhanced Inception(V3) network we proposed finally achieved Analysis and Machine Intelligence, vol. PP, no. 99, pp.
an accuracy of 85.7%. 1–1, 2017.
Today, there are still thousands of families in the world [13] Z. Cui, Y. Liu, and W. Zhao, “Yun: A fast ground-to-
suffering from Alzheimer’s disease. We sincerely hope and air cloud image recognition framework,” in 2019 IEEE
insist that one day, Alzheimer’s disease will be conquered by 23rd International Conference on Computer Supported
human beings, and hope will always be there. Cooperative Work in Design (CSCWD), May 2019, pp.
290–294.
R EFERENCES [14] J. Leng, Y. Liu, T. Zhang, and P. Quan, “Context learn-
[1] K. Blennow, M. J. de Leon, and H. Zetterberg, ing network for object detection,” in ICDM Workshops.
“Alzheimer’s disease,” Lancet, vol. 368, no. 9533, pp. IEEE, 2018, pp. 667–673.
387–403, 2006. [15] J. Leng, Y. Liu, T. Zhang, P. Quan, and
 Z. Cui, “Context-aware u-net for biomedical image learning for image recognition,” 2015.
segmentation,” in IEEE International Conference on [31] Y. Lin, F. Lv, S. Zhu, M. Yang, T. Cour, K. Yu, L. Cao,
Bioinformatics and Biomedicine, BIBM 2018, Madrid, and T. Huang, “Large-scale image classification: fast
Spain, December 3-6, 2018, 2018, pp. 2535–2538. feature extraction and svm training,” in CVPR 2011.
[Online]. Available: http://doi.ieeecomputersociety.org/ IEEE, 2011, pp. 1689–1696.
10.1109/BIBM.2018.8621512 [32] K. Simonyan and A. Zisserman, “Very deep convolu-
[16] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet tional networks for large-scale image recognition,” Com-
classification with deep convolutional neural networks,” puter Science, 2014.
in International Conference on Neural Information Pro-
cessing Systems, 2012.
[17] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wo-
jna, “Rethinking the inception architecture for computer
vision,” in Computer Vision Pattern Recognition, 2016.
[18] N. Otsu, “A threshold selection method from gray-level
histograms,” IEEE Trans.syst.man. Cybern, vol. 9, no. 1,
pp. 62–66, 2007.
[19] S. M. Pizer, E. P. Amburn, J. D. Austin, R. Cromartie,
A. Geselowitz, T. Greer, B. T. H. Romeny, and J. B.
Zimmerman, “Adaptive histogram equalization and its
variations,” Computer Vision Graphics Image Process-
ing, vol. 39, no. 3, pp. 355–368, 1987.
[20] C. R. Jack Jr, M. A. Bernstein, N. C. Fox, P. Thompson,
G. Alexander, D. Harvey, B. Borowski, P. J. Britson,
J. L. Whitwell, C. Ward et al., “The alzheimer’s disease
neuroimaging initiative (adni): Mri methods,” Journal
of Magnetic Resonance Imaging: An Official Journal
of the International Society for Magnetic Resonance in
Medicine, vol. 27, no. 4, pp. 685–691, 2008.
[21] S. Ioffe and C. Szegedy, “Batch normalization: Accelerat-
ing deep network training by reducing internal covariate
shift,” arXiv preprint arXiv:1502.03167, 2015.
[22] P. Jezzard, P. M. Matthews, S. M. Smith et al., Functional
MRI: an introduction to methods. Oxford university
press Oxford, 2001, vol. 61.
[23] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wo-
jna, “Rethinking the inception architecture for computer
vision,” 2015.
[24] J. Leng, Y. Liu, and S. Chen, “Context-aware attention
network for image recognition,” 2019.
[25] J. Mount, “The equivalence of logistic regression and
maximum entropy models,” URL: http://www. win-vector.
com/dfiles/LogisticRegressionMaxEnt. pdf, 2011.
[26] Y. Lcun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-
based learning applied to document recognition,” Pro-
ceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324,
1998.
[27] A. D. N. Initiative, “Access data and samples,” http://
adni.loni.usc.edu/data-samples/access-data/.
[28] ——, “About adni,” http://adni.loni.usc.edu/about/.
[29] A. R. Varma, J. S. Snowden, J. J. Lloyd, P. R. Talbot,
D. M. Mann, and . Neary, D., “Evaluation of the nincds-
adrda criteria in the differentiation of alzheimer’s dis-
ease and frontotemporal dementia,” Journal of Neurology
Neurosurgery Psychiatry, vol. 66, no. 2, pp. 184–8,
1999.
[30] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual

View publication stats