Computer Methods and Programs in Biomedicine 209 (2021) 106294

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

journal homepage: www.elsevier.com/locate/cmpb

A novel multiscale and multipath convolutional neural network based

age-related macular degeneration detection using OCT images
Anju Thomas, P.M. Harikrishnan, Ramachandran, Srikkanth Rajiv Ramachandran,
Rigved Manoj, P. Palanisamy∗, Varun P. Gopi
Department of Electronics and Communication Engineering, National Institute of Technology Tiruchirappalli, Tamilnadu 620015, India

a r t i c l e i n f o a b s t r a c t

Article history: Background and Objective: One of the significant retinal diseases that affected older people is called Age-
Received 23 May 2020 related Macular Degeneration (AMD). The first stage creates a blur effect on vision and later leads to
Accepted 15 July 2021
central vision loss. Most people overlooked the primary stage blurring and converted it into an advanced
stage. There is no proper treatment to cure the disease. So the early detection of AMD is essential to
Keywords: prevent its extension into the advanced stage. This paper proposes a novel deep Convolutional Neural
Age-related macular degeneration Network (CNN) architecture to automate AMD diagnosis early from Optical Coherence Tomographic (OCT)
Multiscale and multipath CNN images.
Classification
Ten-fold cross-validation Methods: The proposed architecture is a multiscale and multipath CNN with six convolutional layers. The
multiscale convolution layer permits the network to produce many local structures with various filter di-
mensions. The multipath feature extraction permits CNN to merge more features regarding the sparse lo-
cal and fine global structures. The performance of the proposed architecture is evaluated through ten-fold
cross-validation methods using different classifiers like support vector machine, multi-layer perceptron,
and random forest.
Results: The proposed CNN with the random forest classifier gives the best classification accuracy results.
The proposed method is tested on data set 1, data set 2, data set 3, data set 4, and achieved an accu-
racy of 0.9666, 0.9897, 0.9974, and 0.9978 respectively, with random forest classifier. Also, we tested the
combination of first three data sets and achieved an accuracy of 0.9902.
Conclusions: An efficient algorithm for detecting AMD from OCT images is proposed based on a multiscale
and multipath CNN architecture. Comparison with other approaches produced results that exhibit the
efficiency of the proposed algorithm in the detection of AMD. The proposed architecture can be applied
in rapid screening of the eye for the early detection of AMD. Due to less complexity and fewer learnable
parameters.
© 2021 Elsevier B.V. All rights reserved.

1. Introduction loss. AMD affects 8.7% of the world’s population, with an estimated
value of 196 million AMD patients by 2020, expanding to 288 mil-
Age-related macular degeneration (AMD) is the most common lion by 2040 [1]. Patients suffering from dry AMD have a signif-
form of the progressive retinal disorder that causes visual im- icant anomaly in the RPE layer known as drusen [2]. The drusen
pairment in developed countries. The early stage of AMD is dry appear as a bulge/perforation. The greater the extent of the prob-
AMD, which produces the abnormality in Retinal Pigment Epithe- lem more significant is the drusen. There is an abnormal growth of
lium (RPE) and generates drusen. AMD’s advanced stage, called the blood vessels under the macula in the wet form, called CNV,
Wet AMD / Choroidal NeoVascularization (CNV), leads to vision which discharges blood and fluid into the retina. Statistics show
that 10% of AMD patients affected by wet form Davis et al. [3].
AMD is the fourth most common cause of blindness in the world
∗
Corresponding author.
[4]. Most people ignore early-stage AMD symptoms and realize the
E-mail addresses: [email protected] (A. Thomas), diseases in their advanced stage, but it was too late to treat them
[email protected] (P.M. Harikrishnan), [email protected] (Ramachandran), and cause blindness. Hence, the identification of AMD in the early
[email protected] (S. Rajiv Ramachandran), [email protected] (R. stage is essential.
Manoj), [email protected] (P. Palanisamy), [email protected] (V.P. Gopi).

https://doi.org/10.1016/j.cmpb.2021.106294
0169-2607/© 2021 Elsevier B.V. All rights reserved.
A. Thomas, P.M. Harikrishnan, Ramachandran et al. Computer Methods and Programs in Biomedicine 209 (2021) 106294

Optical Coherence Tomography (OCT) provides a cross-sectional 3. Materials and method

view of the retina, which worked on the light-wave property. OCT-
based ophthalmology diagnostic method is accepted worldwide 3.1. Materials
due to its non-invasive nature, low cost, and no side effects. The
use of inexpensive and non-radioactive optical contrast agents can The proposed method used four different SD-OCT data sets,
offer safe, highly sensitive, and targeted imaging of regions of in- which are publicly available, more details of data sets are given
terest. Portability, low instrumentation cost, and operational cost below
are other advantages of this mechanism [5]. The cross-section view
of the eye has nine layers; AMD affects one of the layers called
RPE. Typically RPE layer is fat and smooth structured to provide • Data set 1:
central vision acquit, but due to AMD, the RPE layer changes its The data set consists of AMD, DME, and normal image volumes.
texture and produces perforation (dry AMD case), and shifted up- There are 15 image volumes for each category. The OCT data
wards from its flat position (wet AMD). The RPE layer abnormality set acquired by using Spectralis SD-OCT (Heidelberg Engineer-
provides central vision loss in the advanced stage. Hence the de- ing Inc., Heidelberg, Germany) at Duke University, Harvard Uni-
tection of AMD using OCT imaging techniques is very accurate to versity, and the University of Michigan [7]. The AMD category
identify disease by examining the RPE layer. It can be visualized is having dry AMD data set only. The proposed work utilizes 30
with OCT. This paper proposes a novel deep Convolutional Neural image volumes (15 belongs to dry AMD 15 belongs to normal).
Network (CNN) architecture to automate the accurate diagnose of • Data set 2:
AMD in the early stage. The main features of the work are: Data set 2 consists of 50 normal, 48 dry AMD, and 50 DME
OCTs were collected at Noor Eye Hospital in Tehran [15]. The
1. The proposed CNN consist of multiscale and multipath Convo-
axial resolution for this data set is 3.5 μm, with a scan size
lutional Layers (CL)
of 8.9 × 7.4mm2 , although the lateral and azimuthal resolutions
2. The multiscale CL permits the network to produce a large num-
for all patients are not consistent.
ber of local structures with different filter sizes
• Data set 3:
3. The multipath feature extraction permits CNN to merge more
The data set collected from SD-OCT Study, which was registered
features regarding the sparse local and fine global structures
at ClinicalTrials.gov. The data set was collected from National
4. The 16 relevant features from the second fully connected layer
Eye Institute, Devers Eye Institute, Duke Eye Center, and Emory
of the CNN are fed into the classifier reduces the computation
Eye Center [17]. The main criteria were (i) age of the subject be-
time of the classification
tween 50 and 85 years, (ii) drusen height greater than 125 μm
5. Suitable for real-time implementation due to the use of only 6
both eyes or large drusen in one eye with no history of vit-
CL which reduces the complexity and learnable parameters
reoretinal surgery or ophthalmologic disease that might affect
acuity in either eye.
2. Related works • Data set 4:
The data set consist of 108,312 images (37,205 with wet AMD,
Summary of previous works are listed in Tables 1–6. 11,349 with DME, 8,616 with dry AMD, and 51,140 normal) from

Table 1
Existing work in the area of AMD using OCT images.

Performance
Reference Process analysis Data set Highlights Output

Sugruk et al. [6] Noise removal using Gaussian filter, Overall Accuracy: data set 1 [7](10 (i) The testing was carried Dry Normal /
RPE and Retinal Nerve Fiber Layer 0.875 AMD images and 6 out on a limited sample size AMD/ DME
(RNFL) layers are segmented using Diabetic Macular of data (ii) Not working on
thresholding method. Calculated the Edema (DME) highly noised image
mean of RPE value to get baseline of images)
RPE, which will helps to calculate the
height of the drusen. Drusen height
and bubbles in the RNFL layers were
the features in this work.
Srinivasan et al. [7] Histogram of Orientated Gradient Accuracy (AMD - data set 1 [7] (15 Only the volume is labeled Dry AMD/DME/
(HOG) feature was classified using 100%, DME -100%, image volume in as diseased if 33% or more Normal
SVM classifier Normal- 86.67%) each) of the images in the volume
indicate abnormalities
Naz et al. [2] Intensity-based thresholding was used Accuracy: Dry data set 1 [7](20 (i)Baseline estimation get Dry AMD/ Normal
for detecting RPE layer. A curve AMD - 96%, images in each) inaccurate when large
fitting(second order) method was Normal- 100% drusens were present. A
adopted to calculate baseline. Drusen very small portion of the
height as feature. Result verified by original data set was used
using SVM classifier with ten-fold (ii) The work was carried
cross-validation out on a limited sample size
of data
Khalid et al. [8] Selected Region of interest (ROI), RPE Overall data set 1 [7](15 Sensitivity only 56% Dry AMD/Normal
layer extraction using intensity-based Accuracy:92% image volumes in
thresholding, baseline estimation is each)
done by using polynomial fitting.
calculated height of the drusen
(feature)
Karri et al. [9] Classification of image using Overall Accuracy: data set 1 [7] (15 High complexity due to the Dry AMD / DME /
fine-tuned pretrained network 0.940 image volume in use of pre-trained network Normal
(GoogleNet) each)

2
A. Thomas, P.M. Harikrishnan, Ramachandran et al. Computer Methods and Programs in Biomedicine 209 (2021) 106294

Table 2
Existing work in the area of AMD using OCT image.

Performance
Reference Process analysis Data set Highlights Output

Kaymak and Deep learning using AlexNet Overall Accuracy: data set 4 [11] It is highly dependent on Normal/ Dry AMD/
Serener [10] 0.996 the amount of images Wet AMD/ DME
available for training.
Kuwayama et al. To maximize the size of the training Overall Accuracy: Nagoya City (i) Only 100 images were Normal / Wet
[12] set and provide robustness, data 0.850% University included in the test. (ii) Less AMD/ DR ∗ /ERM ∗

augmentation was used. A CNN model Graduate School of accuracy.

was used. Medical Sciences
(1200 images)
Hussain et al. [13] The main features were retinal Overall Accuracy: data set 1 In the case of severe Normal /Dry AMD/
thickness, individual retinal layers 0.977 [7](normal- 59, pathologies, retinal layer DME
thickness, and volume of AMD-177, segmentation failed to locate
hyper-reflective intra-retinal spots and DME-15) them.
drusen. Features were evaluated using
random forest classifier and
performed 15-fold cross validation
Wang et al. [14] Feature was Local Configurable Accuracy: AMD - data set 1 [7](15 The Multiscale Spatial Normal/ Dry AMD/
Patterns (LCP) and opted Sequential 93.3%, image volume in Pyramid (MSSP) was DME
Minimal Optimization (SMO) method Normal-100%, each category) proposed: a spatial pyramid
for classification DME-100% created by scaling and
dividing an image. It’s been
done both with and without
overlaps.
Rasti et al. [15] The image was subdivided with data set 1 AUC- data set 1 [7] data The best classifier was found Dry AMD/DME
different scale, then CNN was used to 0.999, data set 2 set 2 [15] using a Guassian Mixture
classify it. AUC- 0.998, Model(GMM) model. The
precision rate spatial pyramid aided to
(overall) 98.86% avoid over-fitting and reduce
time consumption
∗
Diabetic Retinopathy.
∗
Epiretinal Membranes.

Table 3
Existing work in the area of AMD using OCT image.

Performance
Reference Process analysis Data set Highlights Output

Lee et al. [16] A 21 layered VGG-16 CNN model was Overall Accuracy: Heidelberg Only images of patients that Normal/AMD
used 0.934 Spectralis (Normal met the research
- 48312 images, requirements were used in
AMD- 52690 this study, and the neural
images) network was trained only on
those images. Large number
of layer contribute system
complexity
Farsiu et al. [17] Features: RPE Drusen complex(RPEDC) AUC=0.99 data set 3 [17] To prevent misleading Normal/ AMD
volume thickness, thicking and results, precise retinal layer
thinning volume of abnormal RPEDC, segmentation and manual
and Total Retina(TR) volume. corrections were needed.
performed leave one out method with
linear regression model for
classification
Meng et al. [18] A recombined residual CNN with 18 Accuracy: AMD- data set 1 [7](15 (i) he 18-layer arrangement Normal/Dry
layers 100%, DME- 93.3%, image volumes in creates a complex AMD/DME
Normal -100% each) structure.(ii) Recombined
residual CNN outperformd
the original residual
network.
Yoo et al. [19] Transfer learning using VGG-19 Accuracy (overall): Project Macula Because of the large number Normal /Dry AMD/
pre-trained network was used with 82.6% Data set of parameters, the Wet AMD
random forest classifier pre-trained network with
transfer learning will make
the system more complex.
Sun et al. [20] Features were taken using sparse Accuracy (data set data set 1 [7], Different resolutions of Normal/ Dry AMD/
coding and dictionary learning and 1) Normal-93.33%, private data set input images were used DME
SVM was used for classification AMD- 100%, (AMD-168,
DME-100%, DME-297, and
Accuracy (Private) Normal-213)
Normal-100%, AMD
- 99.67%, DME-
99.67%

3
A. Thomas, P.M. Harikrishnan, Ramachandran et al. Computer Methods and Programs in Biomedicine 209 (2021) 106294

Table 4
Existing work in the area of AMD using OCT image.

Performance
Reference Process analysis Data set Highlights Output

Saha et al. [21] layer segmentation was done by using Accuracy (overall): Private data set Only used private data set hyporefective foci/
ReLayNet and three different 87% hyperrefective foci/
pre-trained networks(ResNet-50, subretinal
Inception-v3 and Inception ResNet-50) drusenoid deposit
were used to find the best one for
AMD detection
Motozawa et al. A nine layer CNN was proposed Accuracy (overall) Private data set limited dataset was used AMD/ Normal
[22] :99%
Yim et al. [23] Introduced Artificial Intelligence (AI) specificity = 90% Private data set (i) Patients’ data from the Wet AMD/ Normal
based system for the detection of previous 11 months was
progressive generation of Wet AMD used to develop a prediction
within 6 months continuous screening algorithm for detecting wet
AMD, (ii)based on the
knowledge made a system
that detect Wet AMD
progression with in 6
months
Xu et al. [24] ResNet-50 was used for the transfer accuracy(overall)- Private datset (i) worked on limited Normal/ Wet AMD/
learning purpose and features were 83.2% dataset (ii) imbalanced daa Dry AMD/
classified using random forest distribution in case of dry polypoidal
classifier and others. (iii) tested data choroidal
set collected from the same vasculopathy(PCV)
device only
Alqudah [25] Generated a new 19- layer CNN accuracy (Overall) data set 4 [11] + Although performed on a Dry AMD/ Wet
network for the multi-classification - 95.30% data set 2 + Private vast number of data sets, AMD/ AMD/ DME/
purpose dataset only 95.3% accuracy was Normal
obtained.
Rong et al. [26] Denoising, thresholding and AUC (Data set 1) = Data set 1 + (i) data set 1 obtained 100% Dry AMD/ Normal
morphologcal dilations were created 0.9856, Private dataset accuracy in volumetric level /DME
for generating mask. After that, the AUC(Private) = calculation
denoised images and masks are used 0.9783
to create a large number of surrogate
images for train the CNN network.

4,686 patients [11]. The proposed method utilizes AMD (dry • Dry AMD:
and wet AMD), and normal images. Dry AMD is a state in which the macula layers get thinner and
drying out; then it leads to loss of functioning of the macula.
The details of the data set used in this work are given in The colour of the macula varies in the early stages of dry AMD.
Table 7. There is a small mass of amorphous material (yellow deposit)
The sample images in the data sets are described below,

Table 5
Existing work in the area of AMD using OCT image.

Performance
Reference Process analysis Data set Highlights Output

Fang et al. [27] In CNN network, current and previous Accuracy (Overall) data set 4 [11] Although performed on a Dry AMD/ Wet
convolution layer features were - 93.4% vast number of datasets, AMD/ DME/
combined iteratively to get better only 93.4% accuracy was Normal
result obtained.
Hwang et al. [28] To build an AI-based system, three Accuracy of data data set 4 [11] + Due to the large number of Normal/ Dry
separate pretrained networks set 4: Inception-v3 Private data set learnable parameter, and AMD/Active wet
(ResNet-50, Inception-v3, and - 96.93%, less accuracy not accurately AMD/ Inactive Wet
VGG-16) were used. ResNet-50 - used for real time AMD
95.87%, VGG-16 - applications
91.20%
Bhatia et al. [29] Pegasus OCT software was used Accuracy (overall) 3 Publc and 3 89% accuracy was achieved General anomaly/
(Based on VGG-16 network as a basic = 89% private dataset after performing on a vast Dry AMD/ Wet
backbone network to create the collection of results. AMD/ AMD/ DME
different classifier networks)
Thomas et al. [30] image denoising based on contrast Accuracy - 96.6% data set 1 On images where the noise Dry AMD/ Normal
enhancement method, based on the pixels intensity in the
information of intensity and position majority of columns are
of RPE layer pixel grouping and more than the RPE and lie
iterative eliminations were performed. below the RPE, it does not
Base line estimation was done with work well.
the help of polymial fitting and
random point selection. Used drusen
height as the feture for classification
Serener and Serte Pretrained ResNet(18 layers) and Accuracy: AlexNet data set 4 [11] The use of pre-trained Dry AMD/ Wet
[31] AlexNet were used - 95.15%, ResNet - networks resulted in a large AMD
99% number of learnable
parameters.

4
A. Thomas, P.M. Harikrishnan, Ramachandran et al. Computer Methods and Programs in Biomedicine 209 (2021) 106294

Table 6
Existing work in the area of AMD using OCT image.

Performance
Reference Process analysis Data set Highlights Output

Huang et al. [32] Retinal layer segmentation was done Accuarcy: 93.96 % Private dataset Less accuracy Dry AMD/ Wet
by using ReLayNet. CNN was applied AMD/DME/Normal
on the extracted layers and finally
fuse it. Classification was baesd on
ten- fold cross validation method
Das et al. [33] multi-scale CNN was proposed Accuracy dataset 4 [34] Using a cost-sensitive loss Dry AMD/ Wet
(overall):97.71% function, the imbalance in AMD/ DME/
the data sets was managed Normal
during learning time.
Kermany et al. [34] Classification based on transfer Accuracy (Overall) data set 4 [11] For classification, Dry AMD/ Wet
learning - 96.53% pre-trained networks were AMD/DME/Normal
used. They are unsuitable for
real-time applications due to
a large number of learnable
parameters and the
network’s complexity.

Table 7
Distribution of Dataset.

Dataset Category No. of subjects /images

data set 1 (Heidelberg SD-OCT: 30 subjects (2 classes: AMD and Normal)) [7] Normal 15
AMD 15
data set 2 (Heidelberg SD-OCT: 98 subjects (Noor Eye Hospital in Tehran consisting of 50 normal, 48 dry AMD)) [15] Normal 50
AMD 48
data set 3 (Bioptigen SD-OCT: 384 subjects (2 classes: AMD and Control)) [17] Normal 115
AMD 269
data set 4 (Mendeley data set: 96961 images (2 classes: AMD and Normal)) [11] Normal 51140
AMD 45821

accumulated in the cells of the eye, which is called drusen, as in the primary path to the first FC layer in the primary path. The
shown in Fig. 1. (b), 1. (d), 1. (f) & 2. (b). The sum of central multipath feature extraction permits CNN to merge more robust
vision loss is proportional to the position of drusen and retinal features regarding the sparse local and fine global structures. The
thinning quantity. multipath compensates for the losses of global structures through
• Wet AMD: an alternate path. As shown in Fig. 3, the alternate route connects
Wet AMD is the irregular blood vessel (which is called CNV) the first CL’s output and makes a bounce to the first FC layer.
generation under the macula. This newly generated blood ves- In the alternate path, sometimes passing the early layer’s feature
sel may leak the fluid and bleed, leading the macula to lift up may misrepresent the CNN to analyze unlinked global structures.
from its flat position as shown in Fig. 2. (c); thus, the central To avoid this issue, an extra CL with a 9×9 filter size with 32 fea-
vision destroys. ture maps is inserted into the alternate path to polish the early
• Normal: layer’s features for improved depictions of the global structure.
In the normal eye, the RPE layer is flat and smooth, and there In the primary path, the first SSCL produces eight feature maps
is no perforation shown in Figures 1. (a), 1. (c), 1. (e) & 2. (a) with a filter size of 7×7. Similarly, the second SSCL having 7×7
filter size to generate 16 feature maps. Next, this is trailed by a
3.2. Proposed method multiscale CL with three different filter sizes is proposed. The mul-
tiscale CL has three branches. The first branch is having a filter size
Traditional image processing with handcrafted features neces- of 3×3. The second branch has a filter size of 5×5 extracts compar-
sitates retinal layer segmentation and feature selection algorithms, atively larger local structures. Similarly, the third branch has a filter
all of which include human expertise. A CNN has many computa- size of 7×7 to cover further local structures. The 96 feature maps
tional layers for extracting features from input images and classi- created by these three branches are combined using depth-wise
fying them into various types, such as the Convolution Layer (CL), concatenation. The 96 primary path feature maps and 32 multipath
Pooling Layer (PL), and Fully Connected (FC) Layer. Due to its suc- feature maps are concatenated and linked to the first FC layer. The
cessful high-level feature extraction, CNN helped to detect AMD in final merged features will describe local and global structures by-
previous studies [35]. This paper aims to propose a new CNN ar- passing the early layer features into the first FC layer. Thus we get
chitecture for AMD classification using OCT images. A multiscale more features, like unique local textures and global features of the
and multipath attribute extraction approach is used in the pro- AMD. After three FC layers, the final output features classify the
posed CNN, which is then classified using various classifiers. OCT images into AMD and normal. The first and second FC layer
Figure 3 depicts the proposed CNN architecture. The proposed contains 128 and 16 hidden neurons, respectively.
CNN uses a 96×96 input image. There are two feed-forward routes Due to the full connectivity structure, the denser FC layer leads
in the proposed CNN. Multiscale CL-based feature extraction is per- to over-fitting. To avoid over-fitting, in the proposed CNN archi-
formed in the primary path. The multiscale neural network-based tecture, a dropout ratio of 0.5 is applied to the first and second
primary path contains two successive Single-Scale Convolutional FC layer [36]. The activation function used to activate the CL is
Layers (SSCL) are accompanied by one multiscale CL and two FC the Rectified Linear Unit (ReLU). The ReLU keeps positive values
layers. The other path is meant for the extraction of the multipath and suppresses negative values to zero. Compared to other activa-
feature. This alternative path is formed by connecting the first SSCL tion functions, ReLU shows better gradient changes [37]. Besides,

5
A. Thomas, P.M. Harikrishnan, Ramachandran et al. Computer Methods and Programs in Biomedicine 209 (2021) 106294

MLP [38] is a nonlinear multi-layer feed-forward network that

maps inputs to outputs. The MLP base structure has three lay-
ers: an input layer, a hidden layer, and an output layer, with
each node entirely linked to the nodes in the next layer with
sufficient weights [39–41].
• Random Forest:
Random forest is an ensemble model classifier that produces a
set of trees with their own random vectors [42,43]. More infor-
mation on the classifier can be found in [39,44].
• Support Vector Machine(SVM):
SVM is a supervised method of learning using related learning
algorithms [45]. Unless the vectors in a space are not linearly
separable, the SVM can help to render them linearly separable
in a higher-dimensional space [46,47].

3.4. Evaluation of performance

The K-fold evaluation is used to assess the classification’s per-

formance. [48,49]. The entire data set available in this assessment
methodology is divided into K-sub parts during the training itself
(K=1,2,3,...). Each sub-section will then be viewed as a validation
set for each iteration. The K value used in this work is 10. The
evaluation results are stored in the form of a confusion matrix
[50]. Table 4 shows the confusion matrix form that represents the
properties of a binary classifier. True positive (TP) and true nega-
tive (TN) indicate accurately classified records, while false positive
(FP) and false negative (FN) indicates information that has been
wrongly classified. We can measure the accuracy, recall, precision,
and F1-score, using these values to assess system performance.
TP + TN
Accuracy = (1)
TP + TN + FP + FN

FP
FalsePositiveRate(FPR ) = (2)
FP + TN

TP
P recision = (3)
TP + FP
Fig. 1. Sample images: (a) & (b) belongs to data set 1 [7], (c) & (d) belongs to data
set 2 [15], (e) & (f) belongs to data set 3 [17]. TP
Recall = (4)
TP + FN
it is easy to implement, and it improves the speed performance of
2 × P recision × Recall
training. After every CL, a Max Pooling Layer (MPL) is employed to F 1 − score = (5)
diminish the spatial resolution of the network’s subsequent layers P recision + Recall
to keep more significant local structures. In addition, the pooling
4. Results and discussion
process increases the size of the receptive field. Because of this,
the network can learn from the input complex local structures.
The proposed CNN training and validation are performed on a
Windows 10 system with Intel i7 @ 3.40 GHz, 64 GB RAM, and
3.3. Classifiers
Nvidia Geforce RTX 2080 11 GB GPU. Back-propagation [51] with
a batch size of 32 is used to train the proposed CNN architecture.
The 16 pertinent features from CNN’s second FC layer are fed
The learning rate α = 0.001 and epochs = 33. Too many epochs
into the following classifiers for classification.
may result in an overfit training dataset, while too few may result
• Multi Layer Perceptron (MLP): in an underfit model. Early stopping is a method that allows to

Fig. 2. Sample images from data set 4 [11].

6
A. Thomas, P.M. Harikrishnan, Ramachandran et al. Computer Methods and Programs in Biomedicine 209 (2021) 106294

Fig. 3. Proposed CNN Architecture.

Table 8
Confusion Matrix for binary classification.

Predicted Class

Normal AMD
Targeted Class Normal TP FN

AMD FP TN

Table 9
Confusion matrix of data set 1 [7] with different classifiers.

Targeted class

Classifier Predicted class Normal AMD

SVM NORMAL 15 0
AMD 2 13
MLP NORMAL 15 0
AMD 3 12
RANDOM FOREST NORMAL 15 0
AMD 1 14

specify an arbitrarily large number of training epochs and then end

training when the model’s performance on the validation dataset
no longer improves. Patience is the number of epochs for which
we evaluate the model’s performance progress before terminating
the training. We assigned a score of ten to patience. In our experi-
ment, the model validation loss was at its lowest at epoch 33 and
did not improve further for the following ten epochs. As a result,
the model training was automatically terminated at epoch 43 with
the optimal model weights at epoch 33 without overfitting. The
criteria we used to save the best model is the validation loss. We
save or checkpoint the best model observed during training as de-
fined by a validation loss on the validation dataset. Whenever the
validation loss reduces from a previous epoch value, we save the
Fig. 4. Proposed model performance plots on Data set 4 [11].
model weight until the performance is not improved further. In our
experiment at epoch number 33, we got the minimum validation
loss with the best model weights. The total number of learnable
parameters of the network is 6,024,512. Table 10
Confusion matrix of data set 2 [15] with different classifiers.

4.1. Proposed method evaluation Targeted class

Classifier Predicted class Normal AMD

The proposed CNN is implemented using the python deep SVM NORMAL 48 2
learning framework. The proposed method is tested using standard AMD 3 45
databases like data set 1 [7], data set 2 [15], data set 3 [17], and MLP NORMAL 48 2
data set 4 [11]. The proposed model performance plot on data set AMD 1 47
RANDOM FOREST NORMAL 50 0
4 is given in Fig. 4. AMD 1 47
The confusion matrices of the data set 1 [7], data set 2 [15],
data set 3 [17], and data set 4 [11] are listed in Tables 9–12 re-

7
A. Thomas, P.M. Harikrishnan, Ramachandran et al. Computer Methods and Programs in Biomedicine 209 (2021) 106294

Table 11
Table 14
Confusion matrix of data set 3 [17] with different classifiers.
Comparison of recent studies on automatic AMD detection.
Targeted class
Weighted average
Classifier Predicted class Normal AMD Method accuracy AUC Data set

SVM NORMAL 114 1 Naz et al. [2] 0.960 Not mentioned

AMD 8 261 Sugruk et al. [6] 0.875 Not mentioned
MLP NORMAL 114 1 Srinivasan et al. [7] 0.933 Not mentioned
AMD 2 267 Khalid et al. [8] 0.920 Not mentioned
RANDOM FOREST NORMAL 115 0 Karri et al. [9] 0.940 Not mentioned data set 1 [7]
AMD 1 268 Hussain et al. [13] 0.977 Not mentioned
Wang et al. [14] 0.966 Not mentioned
Rasti et al. [15] Not mentioned 0.999
Table 12 Meng et al. [18] 1.000 Not mentioned
Confusion matrix of data set 4 [11] with different classifiers. Sun et al. [20] 0.966 Not mentioned
Thomas et al. [30] 0.966 Not mentioned
Targeted class Proposed Method 0.9666 Not mentioned
Classifier Predicted class Normal AMD Rasti et al. [15] Not mentioned 0.998 data set 2 [15]
Proposed Method 0.9897 Not mentioned
SVM NORMAL 48917 2223 Farsiu et al. [17] Not mentioned 0.990 data set 3 [17]
AMD 495 45326 Proposed Method 0.9974 Not mentioned
MLP NORMAL 49329 1811 Proposed Method 0.9902 0.9988 data set 1 [7]+
AMD 393 45428 data set 2
RANDOM FOREST NORMAL 51006 134 [15]+data set 3
AMD 77 45744 [17]
Kaymak and 0.9826 0.9917
Serener [10]
Fang et al. [27] 0.934 0.9798
spectively. The weighted average values of each category are tabu- Hwang et al. [28] 0.9693 0.9835 data set 4 [11]
lated in Table 13. It demonstrates that the performance of the SVM Das et al. [33] 0.9771 0.9900
and MLP classifier is insufficient for the proposed feature extrac- Kermany et al. [34] 0.9653 0.9762
Proposed Method 0.9978 0.9978
tion. The random forest classifier provides less misclassification for
all data sets, which is noticeable. For the random forest classifier,
the average False Positive Rate (FPR) is lower, and the other per-
formance measurements (accuracy, F1-score, recall, and precision)
are all close to one. The proposed CNN feature extraction with ran-
dom forest classifier provides the best result out of other methods.
According to the performance evaluation results, the proposed sys-
tem performs well regardless of the number of training tests, im-
plying that it applies to all data sets. For AMD classification with
the least amount of complexity, the proposed CNN with random
forest classifier appears to be the best choice.
Table 14 shows a comparison of the proposed approach to re-
cent methods. Using 6-CL and ten-fold cross-validation, the pro-
posed method obtained an accuracy of 0.966 for data set 1 [7].
The authors used 15-fold cross-validation to obtain an accuracy of
0.977 in [13]. In the case of severe pathologies, this procedure was
unable to diagnose it. Authors of [18] claims an overall accuracy
of 1, but the result was obtained using 17-CL. Due to 17-CL, this
network was not good in real-time implementation. For data set 2
[15], the authors of [15] achieved an Area Under Curve (AUC) of
0.998, and the proposed method achieved an accuracy of 0.9897. Fig. 5. ROC curves for the various architectures of deep learning using data set 4
[11].
For data set 3 [17], 0.9974 accuracies are obtained using the pro-

Table 13
Weighted average values of the evaluation parameters of different classifiers.

Classifier Dataset FP Rate Precision Recall F1-score Accuracy

SVM 0.067 0.941 0.933 0.933 0.9333

MLP data set 1 [7] 0.100 0.917 0.900 0.899 0.9000
RANDOM FOREST 0.033 0.969 0.967 0.967 0.9666
SVM 0.051 0.949 0.949 0.949 0.9489
MLP data set 2 [15] 0.030 0.970 0.969 0.969 0.9693
RANDOM FOREST 0.011 0.990 0.990 0.990 0.9897
SVM 0.015 0.978 0.977 0.977 0.9765
MLP data set 3 [17] 0.008 0.992 0.992 0.992 0.9921
RANDOM FOREST 0.001 0.997 0.997 0.997 0.9974
SVM 0.021 0.972 0.971 0.971 0.9707
MLP data set 1 [7]+ data set 2 [15]+data set 3 [17] 0.017 0.988 0.988 0.988 0.9882
RANDOM FOREST 0.013 0.990 0.990 0.990 0.9902
SVM 0.026 0.973 0.972 0.972 0.9719
MLP data set 4 [11] 0.021 0.978 0.977 0.977 0.9772
RANDOM FOREST 0.002 0.998 0.998 0.998 0.9978

8
A. Thomas, P.M. Harikrishnan, Ramachandran et al. Computer Methods and Programs in Biomedicine 209 (2021) 106294

Fig. 6. Retinal images (a, c, e) and corresponding Grad-Cam images (b, d, f).

posed method; on the other hand, the authors of [17] achieved has the best performance over other works, as shown by the ROC
an AUC of 0.9900. In the case of the data set 4 [11], it obtained curve.
an accuracy of 99.78%, which is higher than the value among the From Table 14, it is clear that the proposed method is ro-
previous work. Kaymak et al. [10] proposed AlexNet, which was bust in providing noticeable accuracy compared with the exist-
highly sensitive to the training data set. The AlexNet had obtained ing methods. Also, we tested the combination of the 3 data sets
98.26% accuracy and AUC of 0.9917 with 800 epochs. The accu- and achieved an accuracy of 0.9902. From these results, we can
racy of the existing work was less than 98%, and the AUC value see that the proposed method is better if more images are used.
was less than 0.992. Figure 5 shows the ROC curves of all prior The proposed CNN can be applied in real-time due to fewer learn-
studies in AMD detection using data set 4. The suggested work able parameters and less complexity. Since the 16 relevant features

9
A. Thomas, P.M. Harikrishnan, Ramachandran et al. Computer Methods and Programs in Biomedicine 209 (2021) 106294

from the second FCL of CNN are fed into the classifier minimizes References
the computation time of classification. As a result, using CNN as a
feature extractor cuts down on complexity and computation time. [1] W.L. Wong, X. Su, X. Li, C.M.G. Cheung, R. Klein, C.-Y. Cheng, T.Y. Wong, Global
prevalence of age-related macular degeneration and disease burden projec-
This approach compensates for the problem encountered in deep tion for 2020 and 2040: a systematic review and meta-analysis, Lancet Global
learning classification by providing a high recall and precision. Health 2 (2) (2014) e106–e116.
[2] S. Naz, A. Ahmed, M.U. Akram, S.A. Khan, Automated segmentation of RPE
layer for the detection of age macular degeneration using OCT images, in: 2016
Sixth International Conference on Image Processing Theory, Tools and Applica-
4.2. Grad-cam images tions (IPTA), IEEE, 2016, pp. 1–4.
[3] M.D. Davis, R.E. Gangnon, L.Y. Lee, L.D. Hubbard, B. Klein, R. Klein, F.L. Fer-
Another satisfying observation is that the proposed network’s ris, S.B. Bressler, R.C. Milton, The age-related eye disease study severity scale
for age-related macular degeneration: AREDS report no. 17, Arch. Ophthalmol.
activation highlights the RPE layer structure in the OCT images in (Chicago, Ill.: 1960) 123 (11) (2005) 1484–1498.
Fig. 6. The proposed network generates an activation map using [4] T. Vos, R.M. Barber, B. Bell, A. Bertozzi-Villa, S. Biryukov, I. Bolliger, F. Charl-
Gradient-based Class Activation Mapping (Grad-CAM) [52–54] from son, A. Davis, L. Degenhardt, D. Dicker, et al., Global, regional, and national in-
cidence, prevalence, and years lived with disability for 301 acute and chronic
which doctors can directly validate and classify result generated
diseases and injuries in 188 countries, 1990–2013: a systematic analysis for
in cases of misclassification. Grad-CAM finds the gradient of the the global burden of disease study 2013, Lancet 386 (9995) (2015) 743–
last convolutional layer and feature maps concerning the predicted 800.
[5] Z. Yaqoob, J. Wu, E.J. McDowell, X. Heng, C. Yang, Methods and application ar-
class, and then backpropagate the gradient to produce a heat-map
eas of endoscopic optical coherence tomography, J. Biomed. Opt. 11 (6) (2006)
of the same size as the input, when superimposed on the image, 063001.
gives an idea of the discriminative regions of an image. [6] J. Sugruk, S. Kiattisin, A. Leelasantitham, Automated classification between
In class activation map images, the RPE+Bruch’s membrane lay- age-related macular degeneration and diabetic macular edema in OCT image
using image segmentation, in: The 7th 2014 Biomedical Engineering Interna-
ers have very high reflectance and are seen as red pixels. If one tional Conference, IEEE, 2014, pp. 1–4.
considers the activation for dry AMD in Fig. 6 (d) and compares it [7] P.P. Srinivasan, L.A. Kim, P.S. Mettu, S.W. Cousins, G.M. Comer, J.A. Izatt, S. Far-
with the OCT image in Fig. 2 (b), the similarity is evident. The cor- siu, Fully automated detection of diabetic macular edema and dry age-related
macular degeneration from optical coherence tomography images, Biomed.
rugations highlighted in the activation map conform with the cor- Opt. Express 5 (10) (2014) 3568–3577.
rugations present in the RPE+Bruch’s membrane layer later, justify- [8] S. Khalid, M.U. Akram, A. Jameel, T. Khalil, Automated detection of drusens to
ing the decision-making process. On comparing the activation map diagnose age related macular degeneration using OCT images (2017).
[9] S.P.K. Karri, D. Chakraborty, J. Chatterjee, Transfer learning based classification
of wet AMD in Fig. 6 (f) with its OCT in Fig. 2 (c), it can be seen of optical coherence tomography images with diabetic macular edema and dry
that the Bruch’s membrane break into the subretinal pigment ep- age-related macular degeneration, Biomed. Opt. Express 8 (2) (2017) 579–592.
ithelium or subretinal space. For a normal image, the RPE+Bruch’s [10] S. Kaymak, A. Serener, Automated age-related macular degeneration and dia-
betic macular edema detection on OCT images using deep learning, in: 2018
membrane layer is almost a line without any corrugations. In all
IEEE 14th International Conference on Intelligent Computer Communication
three cases, the networks look at the RPE+Bruch’s membrane layer and Processing (ICCP), IEEE, 2018, pp. 265–269.
and decide based on the shape of the RPE+Bruch’s membrane layer. [11] D. Kermany, K. Zhang, M. Goldbaum, et al., Labeled optical coherence tomog-
raphy (OCT) and chest x-ray images for classification, Mendeley Data 2 (2)
(2018).
[12] S. Kuwayama, Y. Ayatsuka, D. Yanagisono, T. Uta, H. Usui, A. Kato, N. Takase,
5. Conclusion Y. Ogura, T. Yasukawa, Automated detection of macular diseases by optical co-
herence tomography and artificial intelligence machine learning of optical co-
herence tomography images, J. Ophthalmol. 2019 (2019).
A novel CNN architecture is proposed to simplify the early de-
[13] M.A. Hussain, A. Bhuiyan, C.D. Luu, R.T. Smith, R.H. Guymer, H. Ishikawa,
tection of AMD. The proposed CNN is a multiscale and multi- J.S. Schuman, K. Ramamohanarao, Classification of healthy and diseased retina
path CNN, where it contains a feed-forward primary path having using SD-OCT imaging and random forest algorithm, PloS one 13 (6) (2018).
multiscale convolution layers and an alternate path. The perfor- [14] Y. Wang, Y. Zhang, Z. Yao, R. Zhao, F. Zhou, Machine learning based detection of
age-related macular degeneration (AMD) and diabetic macular edema (DME)
mance of the proposed architecture is evaluated through ten-fold from optical coherence tomography (OCT) images, Biomed. Opt. Express 7 (12)
cross-validation methods using different classifiers. The proposed (2016) 4928–4940, doi:10.1364/BOE.7.004928.
method is tested on data set 1, data set 2, data set 3, and data set [15] R. Rasti, H. Rabbani, A. Mehridehnavi, F. Hajizadeh, Macular OCT classification
using a multi-scale convolutional neural network ensemble, IEEE Trans. Med.
4 achieved an accuracy of 0.9666, 0.9897, 0.9974, and 0.9978, re- Imaging 37 (4) (2018) 1024–1034.
spectively, with a random forest classifier. Also, we tested the com- [16] C.S. Lee, D.M. Baughman, A.Y. Lee, Deep learning is effective for classifying nor-
bination of the first three data sets and achieved an accuracy of mal versus age-related macular degeneration OCT images, Ophthalmol. Retina
1 (4) (2017) 322–327.
0.9902. As compared to other methods, the findings show that the [17] S. Farsiu, S.J. Chiu, R.V. O’Connell, F.A. Folgar, E. Yuan, J.A. Izatt, C.A. Toth,
proposed algorithm is effective in detecting AMD. The proposed ar- Age-Related Eye Disease Study 2 Ancillary Spectral Domain Optical Coherence
chitecture can be used for fast eye scanning to detect AMD. Tomography Study Group, et al., Quantitative classification of eyes with and
without intermediate age-related macular degeneration using optical coher-
ence tomography, Ophthalmology 121 (1) (2014) 162–172.
[18] T. Meng, C. Wu, T. Jia, Y. Jiang, Z. Jia, Recombined convolutional neural network
Ethical approval statement for recognition of macular disorders in SD-OCT images, in: 2018 37th Chinese
Control Conference (CCC), IEEE, 2018, pp. 9362–9367.
[19] T.K. Yoo, J.Y. Choi, J.G. Seo, B. Ramasubramanian, S. Selvaperumal, D.W. Kim,
This article does not contain any studies with human partici- The possibility of the combination of OCT and fundus images for improving
pants or animals performed by any of the authors. the diagnostic accuracy of deep learning for age-related macular degeneration:
a preliminary experiment, Med. Biol. Eng. Comput. 57 (3) (2019) 677–687.
[20] Y. Sun, S. Li, Z. Sun, Fully automated macular pathology detection in retina op-
tical coherence tomography images using sparse coding and dictionary learn-
Declaration of Competing Interest ing, J. Biomed. Opt. 22 (1) (2017) 016012.
[21] S. Saha, M. Nassisi, M. Wang, S. Lindenberg, S. Sadda, Z.J. Hu, et al., Automated
The authors declare that they have no conflict of interests detection and classification of early AMD biomarkers using deep learning, Sci.
Rep. 9 (1) (2019) 1–9.
[22] N. Motozawa, G. An, S. Takagi, S. Kitahata, M. Mandai, Y. Hirami, H. Yokota,
M. Akiba, A. Tsujikawa, M. Takahashi, et al., Optical coherence tomogra-
Acknowledgment phy-based deep-learning models for classifying normal and age-related macu-
lar degeneration and exudative and non-exudative age-related macular degen-
eration changes, Ophthalmol. Therapy 8 (4) (2019) 527–539.
The authors would like to thank Dr. S Sujatha of Institute of
[23] J. Yim, R. Chopra, T. Spitz, J. Winkens, A. Obika, C. Kelly, H. Askham, M. Lukic,
Ophthalmology, Joseph Eye Hospital, Tiruchirappalli, India for her J. Huemer, K. Fasler, et al., Predicting conversion to wet age-related macular
valuable guidance throughout the research. degeneration using deep learning, Nat. Med. (2020) 1–8.

10
A. Thomas, P.M. Harikrishnan, Ramachandran et al. Computer Methods and Programs in Biomedicine 209 (2021) 106294

[24] Z. Xu, W. Wang, J. Yang, J. Zhao, D. Ding, F. He, D. Chen, Z. Yang, X. Li, W. Yu, [38] Guang-Bin Huang, Yan-Qiu Chen, H.A. Babri, Classification ability of single hid-
et al., Automated diagnoses of age-related macular degeneration and poly- den layer feed forward neural networks, IEEE Trans. Neural Netw. 11 (3) (20 0 0)
poidal choroidal vasculopathy using bi-modal deep convolutional neural net- 799–801, doi:10.1109/72.846750.
works, Br. J. Ophthalmol. (2020). [39] S. Gayathri, A.K. Krishna, V.P. Gopi, P. Palanisamy, Automated binary and mul-
[25] A.M. Alqudah, AOCT-NET: a convolutional network automated classification of ticlass classification of diabetic retinopathy using haralick and multiresolution
multiclass retinal diseases using spectral-domain optical coherence tomogra- features, IEEE Access 8 (2020) 57497–57504.
phy images, Med. Biol. Eng. Comput. 58 (1) (2020) 41–53. [40] P. Manojkumar, C. Surya, P.G. Varun, Identification of ayurvedic medicinal
[26] Y. Rong, D. Xiang, W. Zhu, K. Yu, F. Shi, Z. Fan, X. Chen, Surrogate-assisted plants by image processing of leaf samples, in: Third International Confer-
retinal OCT image classification based on convolutional neural networks, IEEE ence on Research in Computational Intelligence and Communication Networks,
J. Biomed. Health Inf. 23 (1) (2018) 253–263. 2017.
[27] L. Fang, Y. Jin, L. Huang, S. Guo, G. Zhao, X. Chen, Iterative fusion convolutional [41] B. Akbar, V.P. Gopi, V.S. Babu, Colon cancer detection based on structural and
neural networks for classification of optical coherence tomography images, J. statistical pattern recognition, in: 2015 2nd International Conference on Elec-
Vis. Commun. Image Represent. 59 (2019) 327–333. tronics and Communication Systems (ICECS), IEEE, 2015, pp. 1735–1739.
[28] D.-K. Hwang, C.-C. Hsu, K.-J. Chang, D. Chao, C.-H. Sun, Y.-C. Jheng, [42] A. Roychowdhury, S. Banerjee, Random forests in the classification of diabetic
A.A. Yarmishyn, J.-C. Wu, C.-Y. Tsai, M.-L. Wang, et al., Artificial intelli- retinopathy retinal images, in: Advanced Computational and Communication
gence-based decision-making for age-related macular degeneration, Theranos- Paradigms, Springer, 2018, pp. 168–176.
tics 9 (1) (2019) 232. [43] L. Breiman, Random forests, Mach. Learn. 45 (1) (2001) 5–32.
[29] K.K. Bhatia, M.S. Graham, L. Terry, A. Wood, P. Tranos, S. Trikha, N. Jaccard, Dis- [44] S. Gayathri, V.P. Gopi, P. Palanisamy, A lightweight CNN for diabetic retinopa-
ease classification of macular optical coherence tomography scans using deep thy classification from fundus images, Biomed. Signal Process. Control 62
learning software: validation on independent, multicenter data, Retina 40 (8) (2020) 102–115.
(2020) 1549–1557. [45] J.P. Kandhasamy, S. Balamurali, S. Kadry, L.K. Ramasamy, Diagnosis of diabetic
[30] A. Thomas, A. Sunija, R. Manoj, R. Ramachandran, S. Ramachandran, P.G. Varun, retinopathy using multi level set segmentation algorithm with feature extrac-
P. Palanisamy, RPE layer detection and baseline estimation using statistical tion using SVM with selective features, Multimed. Tools Appl. (2019) 1–16.
methods and randomization for classification of AMD from retinal OCT, Com- [46] G. Daqi, Z. Tao, Support vector machine classifiers using RBF kernels with clus-
put. Methods Programs Biomed. 200 (2021) 105822. tering-based centers and widths, in: 2007 International Joint Conference on
[31] A. Serener, S. Serte, Dry and wet age-related macular degeneration classifi- Neural Networks, IEEE, 2007, pp. 2971–2976.
cation using OCT images and deep learning, in: 2019 Scientific Meeting on [47] K. Jemshi, V.P. Gopi, S.I. Niwas, Development of an efficient algorithm for the
Electrical-Electronics & Biomedical Engineering and Computer Science (EBBT), detection of macular edema from optical coherence tomography images, Int. J.
IEEE, 2019, pp. 1–4. Comput. Assisted Radiol. Surg. 13 (9) (2018) 1369–1377.
[32] L. Huang, X. He, L. Fang, H. Rabbani, X. Chen, Automatic classification of retinal [48] S. Yadav, S. Shukla, Analysis of k-fold cross-validation over hold-out validation
optical coherence tomography images with layer guided convolutional neural on colossal datasets for quality classification, in: 2016 IEEE 6th International
network, IEEE Signal Process. Lett. 26 (7) (2019) 1026–1030. Conference on Advanced Computing (IACC), IEEE, 2016, pp. 78–83.
[33] V. Das, S. Dandapat, P.K. Bora, Multi-scale deep feature fusion for automated [49] S. Gayathri, V.P. Gopi, P. Palanisamy, Diabetic retinopathy classification based
classification of macular pathologies from OCT images, Biomed. Signal Process. on multipath CNN and machine learning classifiers, Phys. Eng. Sci. Med. (2021)
Control 54 (2019) 101605. 1–15.
[34] D.S. Kermany, M. Goldbaum, W. Cai, C.C. Valentim, H. Liang, S.L. Baxter, [50] S. Visa, B. Ramsay, A.L. Ralescu, E. Van Der Knaap, Confusion matrix-based fea-
A. McKeown, G. Yang, X. Wu, F. Yan, et al., Identifying medical diagnoses ture selection., MAICS 710 (2011) 120–127.
and treatable diseases by image-based deep learning, Cell 172 (5) (2018) [51] J. Bouvrie, Notes on convolutional neural networks(2006).
1122–1131. [52] R.R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad–
[35] J.H. Tan, S.V. Bhandary, S. Sivaprasad, Y. Hagiwara, A. Bagchi, U. Raghavendra, CAM: visual explanations from deep networks via gradient-based localization,
A.K. Rao, B. Raju, N.S. Shetty, A. Gertych, et al., Age-related macular degenera- in: Proceedings of the IEEE International Conference on Computer Vision, 2017,
tion detection using deep convolutional neural network, Future Gener. Comput. pp. 618–626.
Syst. 87 (2018) 127–135. [53] M. Raju, V.P. Gopi, V. Anitha, K.A. Wahid, Multi-class diagnosis of Alzheimer’s
[36] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout: disease using cascaded three dimensional-convolutional neural network, Phys.
a simple way to prevent neural networks from overfitting, J. Mach. Learn. Res. Eng. Sci. Med. 43 (4) (2020) 1219–1228.
15 (1) (2014) 1929–1958. [54] M. Raju, V.P. Gopi, V. Anitha, Multi-class classification of Alzheimer’s disease
[37] A. Krizhevsky, I. Sutskever, G.E. Hinton, ImageNet classification with deep con- using 3D CNN features and multilayer perceptron, in: 2021 Sixth International
volutional neural networks, in: Advances in Neural Information Processing Sys- Conference on Wireless Communications, Signal Processing and Networking
tems, 2012, pp. 1097–1105. (WiSPNET), IEEE, 2021, pp. 368–373.

11