Instance Selection on CNNs for Alzheimer’s Disease Classiﬁcation from

MRI

J. A. Castro-Silva

1,2,5

, M. N. Moreno-Garc

ıa

, Lorena Guachi-Guachi

3,5

and D. H. Peluffo-Ord

nez

4,5

Universidad de Salamanca, Salamanca, Spain

Universidad SurColombiana, Neiva, Colombia

Department of Mechatronics, Universidad Internacional del Ecuador, Quito, Ecuador

Mohammed VI Polytechnic University, Ben Guerir, Morocco

Smart Data Analysis Systems Group - SDAS Research Group, Ben Guerir, Morocco

Keywords:

Convolutional Neural Network, Instance Selection, Data Leakage, Alzheimer’s Disease.

Abstract:

The selection of more informative instances from a dataset is an important preprocessing step that can be

applied in many classiﬁcation tasks. Since databases are becoming increasingly large, instance selection

techniques have been used to reduce the data to a manageable size. Besides, the use of test data in any part of

the training process, called data leakage, can produce a biased evaluation of classiﬁcation algorithms. In this

context, this work introduces an instance selection methodology to avoid data leakage using an early subject,

volume, and slice dataset split, and a novel percentile-position-analysis method to identify the regions with the

most informative instances. The proposed methodology includes four stages. First, 3D magnetic resonance

images are prepared to extract 2D slices of all subjects and only one volume per subject. Second, the extracted

2D slices are evaluated in a percentile distribution fashion in order to select the most insightful 2D instances.

Third, image preprocessing techniques are used to suppress noisy data, preserving semantic information in the

image. Finally, the selected instances are used to generate the training, validation and test datasets. Preliminary

tests are carried out referring to the OASIS-3 dataset to demonstrate the impact of the number of slices per

subject, the preprocessing techniques, and the instance selection method on the overall performance of CNN-

based classiﬁcation models such as DenseNet121 and EfﬁcientNetB0. The proposed methodology achieved a

competitive overall accuracy at a slice level of about 77.01% in comparison to 76.94% reported by benchmark-

and-recent works conducting experiments on the same dataset and focusing on instance selection approaches.

1 INTRODUCTION

Alzheimer’s disease (AD) is a progressive brain dis-

order and the most common cause of dementia in the

elderly. AD causes nerve cell death and tissue loss

throughout the brain, resulting in a dramatic reduction

in brain volume over time and affecting the major-

ity of its functions. This brain structure is noticeable

on images obtained using various imaging modali-

ties, including Magnetic Resonance Imaging (MRI),

Positron Emission Tomography (PET), and Diffusion

Tensor Imaging (DTI). Due to the increase in life ex-

pectancy and the aging population in developed coun-

tries, it is estimated that AD will affect 60 million

people worldwide over the next 50 years (Ortiz et al.,

2016). There is no cure for AD, and currently avail-

able medications can only help slow the disease’s pro-

gression. As a result, early diagnosis becomes the

best way to ensure effective treatments.

Recently, innovative automatic methods based on

Convolutional Neural Networks (CNNs), which are

part of the deep learning technique, have shown to be

successful in detecting structural changes in the brain

using MRI (Jabason et al., 2019b),(Bae et al., 2020),

(Guan, 2019), (Hussain et al., 2020), (Khan et al.,

2019). CNNs can analyze 2D slices (Farooq et al.,

2017a), (Khan et al., 2019); 3D-patches (Backstrom

et al., 2018), (Zhao et al., 2021); Region-of-interest

(ROI) (Khvostikov et al., 2018), (Lin et al., 2018);

and 3D-subject (Duc et al., 2020), (Backstrom et al.,

2018). Most CNN-based works use a random selec-

tion of training data, which might result in overly op-

timistic or biased models, particularly in cases of data

leakage. Data leakage is often caused by an incor-

rectly split dataset, the lack of an independent dataset

for testing, a late split dataset, or biased transfer learn-

330

Castro-Silva, J., Moreno-García, M., Guachi-Guachi, L. and Peluffo-Ordóñez, D.

Instance Selection on CNNs for Alzheimer’s Disease Classiﬁcation from MRI.

DOI: 10.5220/0010900100003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 330-337

ISBN: 978-989-758-549-4; ISSN: 2184-4313

ing (Wen et al., 2020).

In order to minimize data leakage when devel-

oping CNN-based classiﬁcation models, some works

have proposed different approaches for selecting in-

stances for Alzheimer’s disease classiﬁcation. They

differ in terms of the technique used to obtain the

most representative slices and the number of slices

chosen. For instance, (Farooq et al., 2017b) removes

all slices without informative content. Besides, (Sar-

raf et al., 2016) withdraws the last ten slices as well as

those with a pixel intensity sum of zero. Other auto-

matic techniques (Jabason et al., 2019b), (Khan et al.,

2019) analyze the variation of each slice based on en-

tropy calculation. These techniques select the slices

with the highest entropy values as the most informa-

tive ones. As for the number of slices, works such as

(Hon and Khan, 2017), (Wu et al., 2018), (Qiu et al.,

2018), (Ren et al., 2019), (Khan et al., 2019), (Jaba-

son et al., 2019b) propose using a ﬁxed number of

slices ranging from 30, 32, 48, and 100.

Furthermore, to improve the classiﬁcation per-

formance, researchers apply various image prepro-

cessing techniques such as FreeSurfer (Ren et al.,

2019), (Backstrom et al., 2018) for skull extraction,

segmentation and nonlinear registration; FSL (Duc

et al., 2020), (Zhao et al., 2021) for brain extraction

and tissue segmentation; Statistical parametric map-

ping (SPM) (Farooq et al., 2017a), (Guan, 2019) for

smoothing scans, among others.

Although the data leakage is a problem that af-

fects classiﬁcation models in general, particularly, the

development of solutions that solve the data leak-

age problem by selecting the most informative in-

stances is still a challenging and underexplored task

for AD detection, which demands accurate and un-

biased solutions. Therefore, this work introduces a

methodology for strategically identifying and select-

ing the most informative 2D slices using a percentile-

position-analysis method. The proposed methodol-

ogy intends to reduce data leakage by ensuring that

each subject (patient) belongs to a single subject dis-

tribution set and that only one volume (3D MR im-

age) per subject is selected. Besides, a preprocessing

step is included to use the most informative content of

each 2D slice.

The ability of the proposed methodology to cor-

rectly select the most informative instances is pre-

liminarily explored using two CNN architectures

(DenseNet121 (Huang et al., 2017) and Efﬁcient-

NetB0 (Tan and Le, 2019)), the most well-known in

the state of the art, to classify CN=Normal Cogni-

tion and AD=Alzheimer’s cases from the OASIS-3

dataset.

Furthermore, data leakage behavior is experimen-

tally evaluated by randomly assigning 2D slices to

training, test, and validation sets, which may result in

training data containing information that is intended

to be predicted.

The remaining of this paper is structured as fol-

lows: The materials and methods used for preprocess-

ing and instance selection are included in Section 2.

Section 3 stated the experiment description and pa-

rameter settings of this work. The results and discus-

sion are presented in Section 4. Finally, Section 5

gathers the concluding remarks.

2 MATERIALS AND METHODS

2.1 Dataset

This work uses the OASIS-3 dataset (https://www.

oasis-brains.org/#data), which consists of 3395 T1-

weighted structural magnetic resonance imaging (3D-

MRI) images from 2168 sessions belonging to 1098

subjects ranging in age from 42 to 97 years. Subjects

are characterized using the Clinical Dementia Rating

(CDR) scale, which is a measure that ranges from 0

to 3 often used to determine the overall severity of de-

mentia. A CDR of zero characterizes CN cases, while

a CDR of one or greater characterizes AD cases.

2.2 Proposed Methodology

The proposed methodology aims at selecting the most

informative instances from the dataset to reduce both

the leakage of relevant data and the use of noisy

instances, which could decrease the overall perfor-

mance of a CNN-based classiﬁcation model. The

methodology consists of four stages, as shown in Fig.

1. First, 3D MRI images are prepared to extract 2D

slices of all subjects and one volume per subject. Sec-

ond, the extracted 2D slices are subjected to a novel

percentile-position-analysis (PPA) method in order to

select the most insightful 2D instances. Third, image

preprocessing techniques are used to suppress noisy

data. Finally, the selected instances are used to gener-

ate the training, validation and test datasets.

Data Preparation: This stage starts by randomly

splitting the 3D-MRIs from the OASIS-3 dataset to

ensure that each subject is part of a single subject-

distribution-set (training, validation, or test). This

division guarantees reproducible tests and prevents

data leakage by creating independent training, test,

and validation sets. The subject-distribution-set has

k number of subjects per class, where (k) is less than

Instance Selection on CNNs for Alzheimer’s Disease Classiﬁcation from MRI

331

Figure 1: Workﬂow of the proposed methodology empowered by the novel percentile-position-analysis method for optimal

instance selection.

or equal to the number of samples from the minority

class, to avoid the class imbalance problems. Sub-

sequently, the volume-distribution-set is generated by

selecting one volume for each subject.

Finally, the volume-distribution-set is processed

to extract 2D slices of the orthogonal planes of the

3D MRI (axial, coronal and sagittal). The generated

2D slices are saved in .png format with the image ori-

entation set in RAS (Right-Anterior-Superior).

Instance Selection based on Percentile-Position-

Analysis (PPA) Method: PPA creates 2D-slice

subsets from ﬁve speciﬁc percentile-based positions

-here denoted as P =

{

, p

}

, includ-

ing a ﬁxed number (k) of instances to explore the as-

sociation between slice location and slice content.

For each subset, the initial slice number (i) is com-

puted by equation (1):

i =



100



−

. (1)

where n is the total number of slices per plane, c is the

subset position expressed in percentile, and k is the

desired number of slices. The slice subset (S) includes

the sequence of selected instances from s

to s

i+k

, as

follows

S =

{

i+1

i+2

,··· , s

i+k−2

i+k−1

i+k

}

. (2)

Image Preprocessing: The input image is down-

sampled by standard CNN classiﬁcation models into

smaller images (e.g. 224×224) (Huang et al., 2017),

(Tan and Le, 2019). Down-sampling preserves the se-

mantic information in the image while signiﬁcantly

reducing the number of model parameters. However,

small regions of brain may be vanished from the im-

age by using this technique, making them impossi-

ble to detect its structure. Besides, down-sampling

approaches can reduce the data quality by removing

any essential features that lie at the edges of the im-

age, distorting an image, or adding noisy data. To

address this issue and continue to feed the classiﬁca-

tion model with the most revealing pixels from the 2D

slices, this work examines the following preprocess-

ing techniques:

• Image Trimming: It corrects and standardizes

the brain area by removing black pixel outliers as

depicted in Fig. 2(b).

• Image Resize: Since the size of the 2D slices

varies, the proposed methodology uses three tech-

niques to deﬁne a base size for all slices either

stretching or maintaining the existing aspect ratio,

which is the proportional relationship between an

image width and height:

1. Full Image Resize: The image size is changed

to the new size without preserving the image

aspect ratio, as shown in Fig. 2(f).

2. Resize by Cropping: The longest axis of the

image is cropped to get a square size, and then

it is resized, preserving the aspect ratio as illus-

trated in Fig. 2(e).

3. Resize by Padding: This technique adds black

pixels to the shortest axis to get a square size

image, and then it is changed to the new size,

preserving the aspect ratio. Padding is used to

avoid removing any essential features that lie at

the edges of the image as shown in Fig. 2(d).

• Cropping: It extracts a region of interest (re-

gion of the brain). If the shortest axis is smaller

than the new size, the image is padded to get a

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

332

square image, and then the region of the interest

is cropped as illustrated in Fig. 2(c).

(a) Input image. (b) Trimming. (c) Cropping.

(d) Padding

Resize.

(e) Cropping

Resize.

(f) Full

Resize.

Figure 2: Image preprocessing techniques to suppress the

less informative pixels.

Data Generator: This stage prepares the training

data for batch-by-batch loading since large training

samples often do not ﬁt in memory simultaneously.

The data generator reads the slice metadata dataset

(e.g., slice ﬁlename, class) and preprocesses the im-

ages in training time (without saving to disk) accord-

ing to the desired image output (e.g., image size,

resize technique). The dataset batch-size produced

is based on the computational resources (e.g., GPU,

memory).

3 EXPERIMENTAL SETUP

For experimental tests, 2D slices extracted from 100

CN and 100 AD cases are split into training (60%),

validation (20%), and test (20%) sets before prepro-

cessing. Furthermore, DenseNet121 (Huang et al.,

2017), and EfﬁcientNetB0 (Tan and Le, 2019) are

selected to explore the inﬂuence of the proposed

methodology on the overall performance of a CNN-

based classiﬁcation model. They have been selected

for their successful contribution to the computer vi-

sion ﬁeld as image classiﬁcation, object detection and

localization, scene understanding, and other related

tasks (Goodfellow et al., 2016), (Rosebrock, 2017).

Random search techniques were used to ﬁnd the

optimal hyperparameters using the OASIS-3 dataset

split, as stated before, into three distribution subsets

(train, validation, and test). The proposed methodol-

ogy and all the selected CNNs have been trained with

these hyperparameters capable of achieving the high-

est accuracy to classify Alzheimer’s cases from MRI.

Table 1 provides relevant information about the hy-

perparameter values chosen for the CNN models.

Table 1: Hyperparameter values for DenseNet121 (Huang

et al., 2017) and EfﬁcientNetB0 (Tan and Le, 2019) on

OASIS-3 dataset.

Hyperparameter Value Description

Position 50 Subset position in percentile (1..100).

Anatomical plane Axial Orthogonal plane (Sagittal, Coronal, Axial).

Number of images 32 Number of instances per subject-volume.

Number of channels 3 Number of channels (3=RGB, 1=Gray scale).

Epochs 20 Number of epochs.

Batch size 16 Number of instances by batch.

Transfer learning ImageNet Dataset name.

Optimizer RMSprop Type of optimizer (Adam, SGD, RMSprop).

Learning rate LRS LearningRateSchedule exponential decay.

Initial learning rate 0.0001 A ﬂoat number. The initial learning rate.

Decay steps 10000 A int number. Must be positive.

Decay rate 0.9 A ﬂoat number. The decay rate.

All experiments are reported at the slice level and

were run ﬁve times to obtain consistent results. Addi-

tionally, to obtain a classiﬁcation at the subject level,

all the classiﬁcations obtained at the slice level from

a subject were fused by majority voting. A random

seed was also set for the os, random, TensorFlow, and

NumPy libraries to improve the reproducibility of the

experiments. TensorFlow with Keras and Python li-

braries (including PIL and NumPy) were used to train

and test the explored CNNs with the 2D slices se-

lected through the proposed methodology. Experi-

ments were carried out using a workstation with an

Intel Core i9 9900K processor, 32 GB RAM, and 11

GB NVIDIA RTX 2080Ti GPU.

The inﬂuence of the proposed methodology on

overall CNN performance for Alzheimer’s case clas-

siﬁcation was measured in terms of the average and

standard deviation of accuracy = (T

+ T

)/(T

+ T

+ F

) metric. Where T

, T

, F

, and F

re-

fer to the number of AD cases correctly classiﬁed as

AD, number of CN cases correctly classiﬁed as CN,

number of AD cases misclassiﬁed as CN, and number

of CN cases incorrectly classiﬁed as AD, respectively.

In this sense, accuracy quantiﬁes the proportion of

correctly classiﬁed cases.

4 EXPERIMENTAL RESULTS

AND DISCUSSION

4.1 Data Leakage vs. Independent Data

Sets

This work experimentally evaluates the effect of a

random selection by mixing and shufﬂing all the 2D

slices of all the distribution sets (data leakage), which

might cause training data to contain information that

is intended to be predicted.

Since selecting the most informative slices from

the original data set may improve the overall perfor-

mance of the prediction model (Khan et al., 2019),

Instance Selection on CNNs for Alzheimer’s Disease Classiﬁcation from MRI

333

Table 2: Average accuracy from OASIS-3 with instances from the 50th-percentile (independent dataset) and comparison with

data leakage. Classiﬁcation accuracy measured at subject and slice levels. The highest values are highlighted.

CNN model Slices from the 50th-percentile Data leakage from shufﬂed slices

Subject-Level Slice-Level Subject-Level Slice-Level

DenseNet121 79.00% ± 3.35 75.26% ± 1.35 98.30%± 1.64 96.40% ± 2.66

EfﬁcientNetB0 51.00% ± 1.37 50.96%± 2.76 99.30% ± 0.84 98.08% ± 1.69

Table 3: Average accuracy values achieved by DenseNet121 with different resize techniques. Classiﬁcation accuracy mea-

sured at slice levels. The highest values are highlighted.

Resize Size Planes

Techniques Sagittal Coronal Axial

Full image resize 63.28% ± 2.08 73.87% ± 1.96 76.57% ± 0.93

Resize by cropping 224x224 68.60% ± 1.43 71.02% ± 1.44 75.26% ± 1.35

Resize by padding 60.88% ± 3.41 71.39% ± 0.76 73.95% ± 3.25

Table 4: Average accuracy values achieved by DenseNet121 with different input image sizes. Classiﬁcation accuracy mea-

sured at slice levels. The highest values are highlighted.

Cropping Size Planes

Technique Sagittal Coronal Axial

Cropping Region

64x64 65.15% ± 1.19 70.60% ±2.22 72.07% ± 1.51

96x96 67.39% ± 2.74 72.44% ±1.18 72.04% ± 2.24

128x128 62.21% ± 5.18 70.49% ± 2.76 75.79% ±1.42

Cropping Resize

160x160 67.07% ± 2.61 71.07% ± 3.67 74.04% ±1.17

192x192 70.24% ± 0.87 72.71% ± 2.93 75.90% ±1.14

224x224 68.60% ± 1.43 71.02% ± 1.44 75.26% ±1.35

256x256 71.87% ± 2.21 69.59% ± 2.07 71.42% ±2.75

this experiment also includes results from CNN mod-

els trained with an independent set where the slices of

a subject belong to only one distribution set (training,

validation or test). The dataset was taken from the

50th percentile, as extreme slices often appear black

or have less informative data.

From the results collected in Table 2, it can be

seen that data leakage caused by shufﬂed slices pro-

duces overly optimistic results (higher or equal than

96.40%) when compared to the results of models

trained with the 50th percentile independent dataset

slices (no data leakage). This behavior is repeated for

both subject-level and slice-level classiﬁcation, de-

manding the development of more robust solutions for

selecting more informative slices.

DenseNet121 is the CNN classiﬁcation model

used in the following experiments because it achieves

the highest accuracy (more than 25 %) compared with

EfﬁcientNetB0.

4.2 The Impact of Preprocessing

Techniques

Resizing experiments are conducted on the training

set (in processing time) to evaluate the effectiveness

of each resizing technique on the overall performance

of the CNN-based classiﬁcation model, as shown in

Table 3. All resizing techniques achieve the highest

average accuracy values for the axial planes as they

capture the most critical information of the regions

affected by Alzheimer’s disease.

On the other hand, the average accuracy values

for each technique show that resizing by cropping

technique ensures higher accuracy for all three planes

(sagittal, coronal and axial). Besides, the resizing

by padding technique yields the lowest accuracy due

to the addition of noise when padding the segment.

Since resizing by cropping guarantees a signiﬁcantly

high accuracy value for all three planes, it has been

chosen as the best resizing technique for the proposed

methodology.

Reducing image size leads to information loss.

This experiment examines the impact of image size

on classiﬁcation results. Sizes of 64×64, 96×96

and 128×128 are tried using the cropping region

technique; meanwhile, sizes of 160×160, 192×192,

224×224 and 256×256 are tried using the cropping

resizing technique as shown in Table 4.

From Table 4, it can be observed that the

192×192 size outperforms the average accuracy of

the 224×224 size (experimentally selected for previ-

ous experiments), despite of the fact that almost all

sizes achieved high accuracy for the axial plane.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

334

Table 5: Average accuracy of DenseNet 121 for the selected number of slices. Classiﬁcation accuracy measured at slice levels.

The highest values are highlighted.

Number of instance selected

Plane 1 8 16 32 64

Sagittal 52.50% ±6.01 63.92% ± 1.46 68.21% ± 1.97 68.60% ± 1.43 70.05% ± 2.71

Coronal 50.00% ±6.99 70.85% ± 3.50 68.21% ± 3.29 71.02% ± 1.44 68.38%± 4.11

Axial 54.37% ± 6.48 74.15% ± 1.98 74.23% ± 2.44 75.26%± 1.35 74.98% ± 1.07

Table 6: Average accuracy reached by DenseNet121 which is trained with 2D-slice subsets selected in a percentile distribution

fashion. Classiﬁcation accuracy measured at slice levels. The highest values achieved for each plane are highlighted.

Percentiles explored by the proposed methodology Entropy-based

Plane 20 35 50 65 80 method (Jabason et al., 2019a)

Sagittal 66.63% ±2.46 71.50% ± 3.26 68.60% ±1.43 72.42% ± 2.72 70.73% ±1.49

Coronal 64.14% ±3.55 72.07% ± 1.26 71.02% ± 1.44 70.44% ±0.78 65.44% ± 3.95

Axial 67.68% ± 2.94 77.01% ± 1.61 75.26% ± 1.35 71.95% ±2.06 66.71% ± 1.24 76.94% ± 1.85

4.3 The Impact of the Number of Slices

Selected, and Instance Selection

based on

Percentile-Position-Analysis (PPA)

Method

Number of Slices Selected per Subject: Since the

dataset quantity and quality inﬂuence the ﬁnal perfor-

mance of classiﬁcation models, this experiment eval-

uates the impact of the number of selected informa-

tive slices over classiﬁcation results of DenseNet121

(Huang et al., 2017) by testing subsets of 1, 8, 16, 32,

and 64 slices.

From Table 5, it can be observed that the selec-

tion of 32 2D slices per subject reaches high average

accuracy for planes coronal and axial (71.02% and

75.26%) and 64 slices for the sagittal plane (70.05%).

Furthermore, it is seen that subsets with 64 images

per subject show a slight decrease in accuracy for ax-

ial and coronal planes compared to subsets with 32

slices. This behavior is because adding more im-

age slices with less informative content can result in

redundant, noisy, or less representative information,

lowering CNN performance. On the other hand, sub-

sets with 1 and 8 slices per subject achieve the lowest

accuracy values for all planes. These ﬁndings show

that a number of slices per subject less than or equal

to 8 does not ensure the representativeness of the 170-

256 instances that comprise an MRI volume.

Due to the high accuracy values achieved by sub-

sets with 32 slices, the proposed methodology estab-

lishes the number of 32 slices as the appropriate num-

ber of the most-representative-slices which can be se-

lected for the classiﬁcation of Alzheimer’s cases.

PPA Method: As it is well known, 2D slices from

3D MRI can range from dark to informative images,

and the quality of the content is dependent on the

volume’s position. Therefore, to determine how 2D

slices from different positions and planes affect the

overall performance of DenseNet121 (Huang et al.,

2017) for Alzheimer’s disease classiﬁcation, training

subsets are created with 2D slices distributed in a per-

centile fashion P =

{

, p

}

From results obtained in Table 6, it can be seen

that the most representative 2D slices are located in

the 35th percentile, with the highest accuracy values

for axial (77.01%) and coronal (72.07%) planes. Re-

markably, the most representative 2D slices are found

in the 65th and 35th percentiles for the sagittal plane,

with accuracy values of 72.42% and 71.50%, respec-

tively. This difference indicates a structural symmetry

in the sagittal plane (left and right sides). On the other

hand, it can also be observed that the less informative

slices are found in the extreme percentiles for all three

planes.

Based on the high values of the accuracy average

obtained by the 35th percentile for the axial, sagittal,

and coronal planes, the proposed methodology estab-

lishes the 35th percentile as containing the most infor-

mative instances for the classiﬁcation of Alzheimer’s

disease.

Finally, for comparison purposes, the entropy-

based methodology demonstrated in (Jabason et al.,

2019a) is compared to this work in Table 6. It has

been chosen not only because it is one of the most efﬁ-

cient instance selection techniques with the Densenet

121 architecture, but also because it is similar to the

goal of this work, as it also performs Alzheimer’s

case classiﬁcation. The obtained results show that the

PPA method slightly outperforms the entropy-based

method in terms of overall results. This behavior can

be attributed to the careful assembly of subject and

volume distribution sets, as well as to the optimal se-

lection of the most signiﬁcant instances.

Instance Selection on CNNs for Alzheimer’s Disease Classiﬁcation from MRI

335

5 CONCLUSIONS AND FUTURE

WORK

This work introduces a methodology for strategi-

cally identifying and selecting the most informative

2D slices using a percentile-based-position-analysis

method. The impact of the proposed methodology on

the overall performance of CNN-based classiﬁcation

models is explored experimentally. The slice subsets

contribution to the model performance varies accord-

ing to the position; the 35th percentile reaches the

higher accuracy. Based on the best average results,

the proposed methodology establishes the resize by

cropping technique, the image sizes of (224 × 224)

and (192 × 192) and the axial plane, as suitable to

get the highest model performance for Alzheimer’s

disease classiﬁcation. The number of slices per sub-

ject greatly inﬂuences the model performance, sub-

sets with 32 slices presenting the best results.

The use of 2D slices produces an increased num-

ber of instances and the possibility of using existing

2D CNNs to train a model with transfer learning or

from scratch. The classiﬁcations obtained at the slice

level must be fused to obtain a classiﬁcation at the

subject level. Finally, data leakage can be avoided by

using a subject dataset early split and creating an inde-

pendent test set as proposed in the instance selection

process.

For future work, image metrics will be used to

select the most informative instances. Also, custom

CNNs and model ensembles using the different planes

and cropping regions should be considered to improve

the classiﬁcation model performance and reliability.

ACKNOWLEDGMENTS

This work is supported by the Smart Data Analy-

sis Systems Group - SDAS Research Group (http:

//sdas-group.com)

REFERENCES

Backstrom, K., Nazari, M., Gu, I. Y. H., and Jakola, A. S.

(2018). An efﬁcient 3D deep convolutional network

for Alzheimer’s disease diagnosis using MR images.

Proceedings - International Symposium on Biomedi-

cal Imaging, 2018-April(Isbi):149–153.

Bae, J. B., Lee, S., Jung, W., Park, S., Kim, W., Oh, H., Han,

J. W., Kim, G. E., Kim, J. S., Kim, J. H., and Kim,

K. W. (2020). Identiﬁcation of Alzheimer’s disease

using a convolutional neural network model based on

T1-weighted magnetic resonance imaging. Scientiﬁc

Reports, 10(1):1–10.

Duc, N. T., Ryu, S., Qureshi, M. N. I., Choi, M., Lee, K. H.,

and Lee, B. (2020). 3D-Deep Learning Based Au-

tomatic Diagnosis of Alzheimer’s Disease with Joint

MMSE Prediction Using Resting-State fMRI. Neu-

roinformatics, 18(1):71–86.

Farooq, A., Anwar, S., Awais, M., and Alnowami, M.

(2017a). Artiﬁcial intelligence based smart diagno-

sis of Alzheimer’s disease and mild cognitive impair-

ment. 2017 International Smart Cities Conference,

ISC2 2017, pages 0–3.

Farooq, A., Anwar, S. M., Awais, M., and Rehman, S.

(2017b). A deep CNN based multi-class classiﬁca-

tion of Alzheimer’s disease using MRI. 2017 IEEE In-

ternational Conference on Imaging Systems and Tech-

niques (IST), pages 3–8.

Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep

Learning. MIT Press.

Guan, Z. (2019). A Comprehensive Study of Alzheimer ’ s

Disease. pages 1–16.

Hon, M. and Khan, N. M. (2017). Towards Alzheimer’s

disease classiﬁcation through transfer learning. Pro-

ceedings - 2017 IEEE International Conference on

Bioinformatics and Biomedicine, BIBM 2017, 2017-

Janua:1166–1169.

Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger,

K. Q. (2017). Densely connected convolutional net-

works. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 4700–

4708.

Hussain, E., Hasan, M., Hassan, S. Z., Hassan Azmi, T.,

Rahman, M. A., and Zavid Parvez, M. (2020). Deep

Learning Based Binary Classiﬁcation for Alzheimer’s

Disease Detection using Brain MRI Images. Proceed-

ings of the 15th IEEE Conference on Industrial Elec-

tronics and Applications, ICIEA 2020, pages 1115–

1120.

Jabason, E., Ahmad, M. O., and Swamy, M. N. (2019a).

Classiﬁcation of Alzheimer’s Disease from MRI Data

Using an Ensemble of Hybrid Deep Convolutional

Neural Networks. Midwest Symposium on Circuits

and Systems, 2019-Augus(Mci):481–484.

Jabason, E., Omair Ahmad, M., and Swamy, M. N.

(2019b). Hybrid Feature Fusion Using RNN and Pre-

trained CNN for Classiﬁcation of Alzheimer’s Disease

(Poster). FUSION 2019 - 22nd International Confer-

ence on Information Fusion, pages 2019–2022.

Khan, N. M., Abraham, N., and Hon, M. (2019). Trans-

fer Learning with Intelligent Training Data Selection

for Prediction of Alzheimer’s Disease. IEEE Access,

7:72726–72735.

Khvostikov, A., Aderghal, K., Krylov, A., Catheline,

G., and Benois-Pineau, J. (2018). 3D inception-

based CNN with sMRI and MD-DTI data fusion for

Alzheimer’s disease diagnostics. arXiv.

Lin, W., Tong, T., Gao, Q., Guo, D., Du, X., Yang, Y.,

Guo, G., Xiao, M., Du, M., and Qu, X. (2018). Con-

volutional neural networks-based MRI image analy-

sis for the Alzheimer’s disease prediction from mild

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

336

cognitive impairment. Frontiers in Neuroscience,

12(NOV):1–13.

Ortiz, A., Munilla, J., G

orriz, J. M., and Ram

ırez, J. (2016).

Ensembles of Deep Learning Architectures for the

Early Diagnosis of the Alzheimer’s Disease. Inter-

national Journal of Neural Systems, 26(07):1650025.

Qiu, S., Chang, G. H., Panagia, M., Gopal, D. M., Au, R.,

and Kolachalama, V. B. (2018). Fusion of deep learn-

ing models of MRI scans, Mini–Mental State Exami-

nation, and logical memory test enhances diagnosis of

mild cognitive impairment. Alzheimer’s and Demen-

tia: Diagnosis, Assessment and Disease Monitoring,

10(September):737–749.

Ren, F., Yang, C., Qiu, Q., Zeng, N., Cai, C., Hou, C., and

Zou, Q. (2019). Exploiting Discriminative Regions of

Brain Slices Based on 2D CNNs for Alzheimer’s Dis-

ease Classiﬁcation. IEEE Access, 7:181423–181433.

Rosebrock, A. (2017). Deep Learning for Computer Vision

with Python:StarterBundle. Deep learning for com-

puter vision with Python. PyImageSearch.

Sarraf, S., DeSouza, D., Anderson, J., and Toﬁghi, G.

(2016). DeepAD: Alzheimer’s Disease Classiﬁcation

via Deep Convolutional Neural Networks using MRI

and fMRI. bioRxiv, page 070441.

Tan, M. and Le, Q. (2019). Efﬁcientnet: Rethinking model

scaling for convolutional neural networks. In Interna-

tional Conference on Machine Learning, pages 6105–

6114. PMLR.

Wen, J., Thibeau-Sutre, E., Diaz-Melo, M., Samper-

Gonz

alez, J., Routier, A., Bottani, S., Dormont, D.,

Durrleman, S., Burgos, N., and Colliot, O. (2020).

Overview of classiﬁcation of Alzheimer’s disease.

Medical Image Analysis, 63.

Wu, C., Guo, S., Hong, Y., Xiao, B., Wu, Y., and Zhang, Q.

(2018). Discrimination and conversion prediction of

mild cognitive impairment using convolutional neu-

ral networks. Quantitative Imaging in Medicine and

Surgery, 8(10):992–1003.

Zhao, Y., Ma, B., Jiang, P., Zeng, D., Wang, X., and Li, S.

(2021). Prediction of Alzheimer’s Disease Progres-

sion with Multi-Information Generative Adversarial

Network. IEEE Journal of Biomedical and Health In-

formatics, 25(3):711–719.

Instance Selection on CNNs for Alzheimer’s Disease Classiﬁcation from MRI

337