Computer Methods in Biomechanics and Biomedical

Engineering: Imaging & Visualization

ISSN: (Print) (Online) Journal homepage: www.tandfonline.com/journals/tciv20

Early detection of glaucoma: feature visualization

with a deep convolutional network

Jisy N K, Md. Hasnat Ali, Sirisha Senthil & M.B. Srinivas

To cite this article: Jisy N K, Md. Hasnat Ali, Sirisha Senthil & M.B. Srinivas (2024) Early
detection of glaucoma: feature visualization with a deep convolutional network, Computer
Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 12:1, 2350508,
DOI: 10.1080/21681163.2024.2350508

To link to this article: https://doi.org/10.1080/21681163.2024.2350508

© 2024 The Author(s). Published by Informa

UK Limited, trading as Taylor & Francis
Group.

Published online: 20 May 2024.

Submit your article to this journal

Article views: 69

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tciv20
COMPUTER METHODS IN BIOMECHANICS AND BIOMEDICAL ENGINEERING: IMAGING & VISUALIZATION
2024, VOL. 12, NO. 1, 2350508
https://doi.org/10.1080/21681163.2024.2350508

Early detection of glaucoma: feature visualization with a deep convolutional network

a
Jisy N K , Md. Hasnat Alia,b, Sirisha Senthil b
and M.B. Srinivasc
a
Electrical and Electronics Engineering, BITS Pilani, Hyderabad Campus, Hyderabad, India; bLV Prasad Eye Institute, Hyderabad, India; cElectrical and
Electronics Engineering, BITS Pilani, Dubai Campus, Dubai, UAE

ABSTRACT ARTICLE HISTORY

Computer-aided diagnosis of ocular disorders like glaucoma can be effectively performed with retinal Received 16 May 2023
fundus images. Since the advent of machine learning and later deep learning techniques in medical Accepted 27 April 2024
image analysis, research focusing on automated and early detection of glaucoma has gained promi­ KEYWORDS
nence. In this paper, we show how a deep Convolutional Neural Network model can be used in visualising Glaucoma; fundus images;
the features related to Glaucoma that are consistent with (a subset of) those used by ophthalmologists for convolutional neural
diagnosis. However, since the doctors do not necessarily depend entirely on these features but back them network; deep learning;
up with other important measurements such as visual field testing (HVF), any deep neural network such visualization
as CNN may not necessarily be expected to yield 100% accuracy. Further, other important parameters
such as Intraocular Pressure (IOP), which may be built into a deep learning model, may not necessarily
correlate well with the presence of Glaucoma. In this work, a deep CNN model has been trained and
tested on a large number of high-quality fundus images, both normal and glaucomatous. We compare
the results with transfer learning models such as VGG16, ResNet50, and MobileNetV2. On an average, we
obtained an accuracy of 93.75% in identifying glaucoma, focusing only on the features of the fundus
image.

Introduction The Optic Disc (OD) has certain features, under healthy and
damaged conditions, that can be identified and learned from
Glaucoma, an ocular condition that leads to irreversible blind­
the fundus images. Figure 1a below shows the normal fundus
ness affecting millions of people across the world, is a serious
image while Figure 1b represents a glaucomatous optic disc,
cause for concern. Early detection of this condition and appro­ and Figure 1c is a glaucomatous disc image with optic disc and
priate treatment may certainly help halt the progression and optic cup boundaries marked by an expert. Glaucoma diagnosis
prevent blindness. Glaucoma is a sight-threatening ocular con­ is typically performed based on the physical evaluation of optic
dition characterised by progressive loss of retinal ganglion cells nerve head (ONH) via ophthalmoscopy, the patient’s visual field
(RGC) (Dimitriou and Broadway 2013). Typical optic nerve tests (perimetry), and additional clinical information such as
changes in glaucoma include an increase in the cupping of Intraocular Pressure (IOP), etc (Bock et al. 2010). Clinical evalua­
the optic nerve head (ONH), neuro-retinal rim thinning, and tion of Optic Nerve Head (ONH) is the cornerstone of diagnos­
loss of retinal nerve fibres. While advanced glaucoma may not ing glaucoma (Hussain and Holambe 2015). Further, thinning of
be difficult to diagnose by a general ophthalmologist, the Retinal Nerve Fiber Layer (RNFL) is also an indicator of glau­
challenge is with early detection of this blinding condition coma which while is clearly visible in Optical Coherence
which would require a trained glaucoma specialist. Tomography (OCT) images, is only faintly visible in fundus
Retinal imaging techniques have been developed in past images. The risk factors associated with glaucoma include
years to identify patients with retinal as well as systemic dis­ age, gender, height, weight, family history with glaucoma
eases. Fundus imaging methods are adopted to diagnose ret­ (FHG), etc. and systemic diseases like diabetic Mellitus and
inal damages caused by pathologies such as diabetic hypertension (Coleman and Miglior 2008; McMonnies 2017;
retinopathy (DR), age-related macular degeneration (AMD), Alam Miah and Yousuf 2018).
and glaucoma (Johora et al. 2020). In medicine and healthcare,
the use of fundus photographs is considered a prominent
Feature extraction from fundus images
method to detect these ocular diseases. Fundus images provide
crucial information such as optic disc area, optic cup area, rim For feature extraction from fundus images, several image
area, cup-to-disc ratio (CDR), blood vessel bending, and retinal processing techniques have been implemented (Khan et al.
nerve fibre layer (RNFL) thickness that helps to identify glau­ 2019; Saiteja et al. 2021). Emergence of AI has enabled
coma. Also, for large-scale screening and diagnosis of glau­ researchers to use these features in conjunction with machine
coma, using fundus images is considered a popular approach learning models to help improve glaucoma diagnosis.
(Bernardes et al. 2011). Machine learning (ML) allows computers to learn from

CONTACT Md. Hasnat Ali [email protected] Electrical and Electronics Engineering, BITS Pilani, Hyderabad Campus, LV Prasad Eye Institute, Kallam Anji
Reddy campus, LV Prasad Marg, Banjara Hills, Hyderabad 500034, India
© 2024 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License (http://creativecommons.org/licenses/by-nc/4.0/), which permits
unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited. The terms on which this article has been published allow the
posting of the Accepted Manuscript in a repository by the author(s) or with their consent.
2 J. N K ET AL.

Figure 1. Fundus image (a) healthy eye (b) glaucomatous eye (c) glaucomatous image marked by the expert.

datasets (fundus images in the present case) and make effec­ validate the same with those extracted through visualisa­
tive predictions (Sarhan et al. 2020). While ML models give tion by CNN layers.
the best performances with small datasets, due to manual ● The proposed model is compared with ImageNet trained
feature selection from fundus images, the convergence and CNN architectures such as VGG16, ResNet50 and
over-fitting problems in the models are not eliminated. This MobileNetV2, and it has been observed that our model
limits their application and thus use of deep learning (DL) outperformed these architectures yielding an accuracy of
models has become popular due to their robust diagnostic 93.75 %.
performance in ophthalmic disease classification (Lecun et al.
2015; Bulut et al. 2020). Convolution Neural Networks (CNNs) The paper is organised as follows: Section 2 summarises the
are DL architectures that can be effectively implemented for recent literature related to the application of deep learning
medical image segmentation and classification as they can techniques for glaucoma classification. Section 3 provides details
extract features and learn from the pixel intensities of the of the deep CNN model deployed by us that has built-in visua­
images (Diaz-Pinto et al. 2019). It may however be noted that lisation features. Section 4 describes the ImageNet trained CNN
while ML models based on radiomic features, generated by architectures and compares the results obtained with those of
the extraction of quantitative features such as, shape fea­ our model. Section 5 presents the experimental results obtained
tures, first-order histogram features, and second-order texture with visualisation of features as well as the performance metrics
features, etc., provide better explainability (Leandrou et al. while Section 6 draws conclusions of the work.
2023; Militello et al. 2023; Prinzi et al. 2023), DL architectures
typically have better predictive performance in medical
image analysis (Sun et al. 2020; Wei 2021).
Literature survey
The deep CNN model uses more convolution layers to
extract features from the raw pixel image data. It can make Out of all the deep learning approaches for image classification,
automated predictions from the input image data using built-in CNN is the most preferred one since it can learn and extract
feature extraction property. Since the model should extract all highly discriminative features from pixel intensities of the
salient features from the fundus images, localisation of the image. CNN is used by researchers for medical image analysis
region of interest in the images for better feature selection is and classification and thus it’s no surprise that it is also used for
the preliminary step before resorting to the deep learning glaucoma detection and classification (Elangovan and Nath
approach. Unlike the approach in current literature (Orlando 2021; Alayón et al. 2023). A deep learning approach was devel­
et al. 2017; Diaz-Pinto et al. 2019; Park et al. 2020) that uses oped to detect optic disc abnormality and to classify optic disc
fundus images cropped around the OD or entire fundus image, regions into three classes using four public and four private
our work proposes to use images cropped around ONH, which databases (Alghamdi et al. 2017). The three classes of glaucoma
is at the centre of the image as well as the RNFL region that condition used in their approach were normal, suspicious, and
surrounds it even if it appears faint. Further, visualisation by abnormal classes. In their work, the optic disc localisation
deep network layers has been incorporated to identify the approach used CNN to learn the most informative features
features extracted between the layers to understand the fea­ from the fundus image. But CNN requires a large amount of
tures being looked into by the model. labelled data. As a result, training deep CNN from scratch is
The main contributions of this work are as follows. difficult. However, this was surpassed by using pre-trained CNN
with adequate fine-tuning (Tajbakhsh et al. 2017). Two different
● For better feature selection, we pre-processed the fundus CNNs were deployed to produce the feature vectors (Orlando
images with a region of interest (ROI) that incorporated et al. 2017). The CNNs used in their work were VGG-S and
the ONH and RNFL region. Experiments have been per­ OverFeat. Pre-processing techniques such as vessel inpainting,
formed on these images for glaucoma classification. contrast limited adaptive histogram equalisation (CLAHE), and
● We propose a version of deep CNN model to predict the cropping around ONH were proposed. However, the main lim­
stage of glaucoma (mild, moderate, and severe) based on itation of their work includes a small number of image datasets
features typically used by the ophthalmologists and to test the performance of CNN.
 COMPUTER METHODS IN BIOMECHANICS AND BIOMEDICAL ENGINEERING: IMAGING & VISUALIZATION 3

A Glaucoma-Deep system was developed that implemented databases which produced better results than CDR-based glau­
an unsupervised CNN architecture which automatically extracts coma classification. Segmentation and classification of glaucoma
features through a multilayer from raw pixel intensities of was implemented using U-Net architecture with deep CNN model
colour fundus images (Abbas 2017). Here, a deep-belief net­ in the work by (Sudhan et al. 2022). The model used ORIGA dataset
work (DBN) model was used to select the most discriminative and deep CNN (DCNN) approach was used for classification of
features from the CNN model. In their model, retinal images glaucoma. In all the research work reported above, deep learning
from three public and one private database were used to approach was used on public or private datasets to extract features
evaluate the performance using statistical measures like sensi­ in an attempt for efficient classification. It appears however that
tivity (SE), accuracy (ACC), specificity (SP), and precision (PRC). feature visualisation from fundus images for glaucoma detection
For automatic glaucoma assessment using retinal fundus has not been attempted.
images, five different ImageNet trained models were employed A deep learning model was investigated to identify the
(Diaz-Pinto et al. 2019). Extensive validation of CNNs was done in best X-ray features for COVID-19 detection in the work by
their work. Images were taken from ACRIMA, the largest public (Apostolopoulos et al. 2022). The grad-CAM algorithm was
database for glaucoma diagnosis, and out of the ImageNet trained used to visualise the areas of the images in their work. To
models (InceptionV3, VGG16, VGG19, Xception, and ResNet50), the efficiently interpret and diagnose glaucoma, an explainable
Xception architecture resulted in the best performance. CDR machine learning model was developed in the work by (Oh
detection and ONH localisation was developed using state-of-the- et al. 2021). They designed a system called Magellan that
art DL architectures such as YOLO V3, ResNet, and DenseNet to produced explainable results in the prediction of glau­
compare various aspects of performance other than the classifica­ coma. Glaucoma prediction analysis was evaluated using
tion accuracy (Park et al. 2020). The performance measures dis­ an explainable artificial intelligence model in the work by
cussed in their work included processing time, image resolution, (Kamal et al. 2022). The author proposed explainable arti­
and diagnostic performance. An ensemble of deep convolutional ficial intelligence and interpretable machine learning
networks using stacking ensemble learning technique was devel­ model to generate trustworthy predictions for glaucoma.
oped for glaucoma detection (Elangovan and Nath 2022). Their However, in the above works, feature visualisation appears
approach was implemented with thirteen pre-trained models such to have played a lesser role in glaucoma detection and
as Alexnet, VGG-16, VGG-19, Googlenet, Resnet-18, Resnet-50, classification. A major focus of this work is the visualisation
Resnet-101, Squeezenet, Mobilenet-v2, Efficientnet-b0, Inception- of features between layers of our deep CNN model and
v3, Xception and Densenet-201 on modified publicly available utilisation of the same for glaucoma classification.
databases such as DRISHTI-GS1-R, ORIGA-R, RIM-ONE2-R, LAG-R,
and ACRIMA-R. A DL approach was developed for pathological
Methodology
area localisation and diagnosis of glaucoma (Li et al. 2020). An
attention-based CNN (AG-CNN) using large-scale attention-based Our approach is to take advantage of the visualisation ability of
glaucoma (LAG) database for glaucoma classification was estab­ deep CNN layers while assessing the features of fundus images.
lished, and further, performance metrics were also evaluated. Optic Accordingly, the whole dataset is pre-processed to the required
disc and optic cup segmentation was performed using deep learn­ size and resolution. Also, the dataset is augmented to increase
ing convolutional network in the work by the authors (Juneja et al. the number of images. These images are fed to the deep CNN
2020). The authors used small dataset and a deep learning archi­ architecture with four convolutional layers and two dense
tecture was developed with CNN for automating glaucoma detec­ layers for the classification of glaucoma and normal images.
tion. Statistical parameters such as cup entropy and kurtosis from Visualisation of the features is carried out in between the layers
colour fundus images were measured in the work by (Elangovan which gives us clarity on the features extracted from within the
et al. 2020) using FCM algorithm based on morphological recon­ layers. The proposed methodology is depicted in Figure 2 while
struction and membership filtering (FRFCM). They implemented details of pre-processing of the fundus dataset and deep CNN
the algorithm on publicly available RIM-ONE and DRIONS-DB architecture are provided in the subsections below.

Figure 2. Proposed deep CNN model with visualization of features.

4 J. N K ET AL.

Dataset Preprocessing of fundus dataset

The fundus dataset considered in this work consists of 295 Selection of region of interest
normal images and 120 glaucomatous images which are The region of interest (ROI) from fundus images is chosen in
part of a project carried out by L.V. Prasad Eye Institute, such a way that important features related to glaucoma are
based in Hyderabad, India. Normal fundus images are retained. Pre-processing of images was carried out using
divided, based on the optic disc size, into large, medium Python and the OpenCV library. The first step in preprocessing
and small with a count of 96, 100, and 99, respectively. is to remove the redundant portions of the fundus image
Similarly, optic discs were divided into small (<1.6 mm2), followed by cropping around OD, cropping around ONH
average (1.6–2.6 mm2) and large (>2.6 mm2) (Rao et al. while retaining the RNFL region, segmented optic disc, as well
2009). Glaucomatous images have been divided into mild, as using the full fundus image in its entirety. These methods are
moderate and severe based on the severity of damage illustrated in Figure 3 for a glaucomatous fundus image.
(based on visual field severity grading done by trained The features relevant for glaucoma diagnosis are those that
glaucoma specialist). Severity of glaucoma is assessed surround the optic disc and the peripapillary (around the optic
based on visual fields mean deviation (MD), viz., mild: MD disc) RNFL. Thus, out of the four types of pre-processed images,
better than −6 dB, moderate: MD between −6 dB and −12 cropping around ONH that includes RNFL region, has been
dB and severe: MD worse than −12 dB, known as Hoddap- chosen for feature extraction and classification as explained in
Parrish-Anderson criteria (E et al. 1993). They are of a total (Jisy et al. 2024). Subsequently, the fundus images have been
count of 10, 23, and 29, respectively. The original image resized to 256 × 256 and labelled appropriately before being
resolution is 3216 × 2136 which is very large for any deep given as input to our deep CNN model. A sample of the cropped
learning model to handle. Hence, we have pre-processed and resized images is shown in Figure 4 below. This downscaled
the entire image dataset to a manageable size. resolution helps to reduce the time complexity of computation.

Figure 3. Pre-processing of fundus image with glaucoma (a) entire image (b) fundus image cropped around OD (c) fundus glaucoma image cropped around ONH
including RNFL region (d) segmented optic disc region.

Figure 4. Examples of the ROI in the fundus images cropped around the ONH including RNFL region for both normal and glaucomatous images.
 COMPUTER METHODS IN BIOMECHANICS AND BIOMEDICAL ENGINEERING: IMAGING & VISUALIZATION 5

Table 1. Fundus dataset before and after augmentation. this layer is connected to all the neurons in the previous layers.
Fundus glaucoma class Total images Images after augmentation While training the network, the loss function used is categorical
Mild 10 50 cross entropy with RMSprop optimiser. The learning rate is 0.0001
Moderate 23 118
which helps to achieve the minimum cost function. The network
Severe 29 137
was trained for 100 epochs with a batch size of 16. The steps per
epoch while training the model are chosen according to the batch
Data augmentation size. All these hyperparameters are effectively selected, and the
Since the dataset is of a smaller size, we performed data aug­ weights or system parameters are updated and saved appropri­
mentation to increase the number of images for training and ately. The details of kernel size and filter size are mentioned in
testing the model. Data augmentation can be performed by Table 2. Also, visualisation of features within the layers was
several transformations on the image or by synthetically gen­ resorted to and observed accordingly.
erating images using techniques such as generative adversarial
networks (GAN) (Nandhini Abirami et al. 2021). We have aug­ Performance metrics
mented the dataset by image transformations like flipping
horizontally, flipping vertically, and rotating the image by 90°, The results are evaluated by computing performance measures
180°, and 270°. These techniques are applied to all the classes of such as accuracy, specificity, sensitivity, precision, F1 score and
glaucoma image datasets. In our architecture dropout layers Matthews Correlation Coefficient (MCC). The true positive (TP),
and normalisation terms are also added. The number of images false positive (FP), true negative (TN), and false negative (FN)
obtained after augmentation is presented in Table 1 below. values are identified to define the performance metrics. TP is
glaucoma image being identified as glaucoma, TN is the
healthy image being identified as healthy, FP is a false glau­
The deep CNN architecture coma prediction which means normal is identified as glaucoma
and FN is a false normal prediction in which glaucoma is
The proposed architecture uses four convolutional layers (four
identified as normal. The equations for evaluating the accuracy,
convolution layers + four max pooling layers) and two fully con­
specificity, sensitivity, precision, F1 score and MCC are given
nected layers. This is depicted in Figure 5. According to the input
below (Bogacsovics et al. 2022; Anbalagan et al. 2023)
image dataset, the number of layers and filter sizes are selected
appropriately such that the number of parameters involved in TP þ TN
Accuracy ¼ (1)
the design is also optimised. The input image size is 256 × 256 × TP þ FN þ TN þ FP
3 representing the image dimensions, height x width x channel.
We chose the ReLU activation function for all convolution
TP
layers as it helps in vanishing gradient problems in which useful Sensitivity ¼ (2)
TP þ FN
gradient information is unable to propagate from output to input.
Dropout layers are also implemented in between the layers to
reduce the overfitting problem. The fully connected output layer TN
uses the softmax activation function. More number of parameters Specifity ¼ (3)
TN þ FP
are involved in the implementation of this layer as each neuron in

Figure 5. The proposed deep CNN architecture.

6 J. N K ET AL.

Table 2. Details of the proposed deep CNN architecture.

Layers Kernels Filter size Output shape
Conv2D 1 32 3x3 254 x 254 x 32
MaxPool2D 1 32 2x2 127 x 127 x 32
Conv2D 2 64 3x3 125 x 125 x 64
MaxPool2D 2 64 2x2 62 x 62 x 64
Conv2D 3 128 3x3 60 x 60 x 128
MaxPool2D 3 128 2x2 30 x 30 x 128
Conv2D 4 256 3x3 28 x 28 x 256
MaxPool2D 4 256 2x2 14 x 14 x 256
Flatten – – 50176
Fully connected Dense1 256 – 256
Fully connected Dense2 2 – 2

looking into the depth of the layers and their performances in

TP
P recision ¼ (4) medical image analysis. VGG16 architecture is 16 layers deep
TP þ FP (13 convolutional layers + 5 maxpooling layers + 3 dense layers)
that contributes to the learnable parameters in the network
(Simonyan and Zisserman 2015). The convolution layer is a 3 ×
2�precision�sensitivity 2�T P
F 1score ¼ ¼ 3 filter with stride 1 that uses same padding while maxpool
precision þ sensitivity 2�T P þ F P þ F N
layer is 2 × 2 filter with stride 2. ResNet50 architecture is 50
(5)
layers deep that comprises of convolutional blocks and identity
blocks. This architecture reduces the vanishing gradient pro­
blem encountered with VGG16 networks and also the number
T P :T N F P :F N
MCC ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi of trainable parameters is less than VGG16 model (Khan et al.
ðT P þ F P Þ:ðT P þ F N Þ:ðT N þ F P Þ:ðT N þ F N Þ
2020). MobileNetV2 is a CNN architecture which is 53 layers
(6) deep. The architecture includes depth wise separable convolu­
tions with linear bottlenecks. Information within convolution
layers is retained in the inverted residual blocks (Sandler et al.
2018). The model performed well for image classification (Dong
ImageNet trained CNN architectures et al. 2020). All the models help to extract features and classify
the image dataset using transfer learning (TL) approach.
CNN architecture has evolved from a few layers to multiple
layers in the past few years to extract high level features of
the application on hand. ImageNet is a large database of
Transfer learning
images that comprises of millions of annotated images segre­
gated into different classes. Making use of this database in the Transfer learning (TL) is a technique of transferring the knowl­
annual ImageNet Large Scale Visual Recognition Challenge edge learned from training a model in one domain to
(ILSVRC), the performance of ImageNet trained CNN architec­ a different domain. The local features such as edges and
tures has been evaluated (Russakovsky et al. 2015). Such corners of the images can be captured by this method using
ImageNet trained models have been used in glaucoma disease pre-trained weights obtained from the architectures (Yosinski
classification (Diaz-Pinto et al. 2019; Park et al. 2020; et al. 2014). This helps to reduce the training time and data
Bogacsovics et al. 2022). But the results and model perfor­ required. The VGG16, ResNet50 and MobileNetV2 architec­
mance vary according to the specific features selected and tures with pre-trained weights are used to extract features.
the datasets used for training. In this work, we compared the Fine tuning these models as a classifier generating softmax
results obtained for different glaucoma classes with three probabilities is carried out to reduce the number of trainable
ImageNet trained architectures, viz., VGG16, ResNet50 and parameters. The process of training and updating the pre-
MobileNetV2. The models were selected for comparison by trained weights of CNN using backpropagation is called fine-

Table 3. The list of hyperparameters for the ImageNet trained architectures.

Type of glaucoma Network Optimizer Batch size Learning rate Epochs
Mild VGG16 Adam 16 0.001 100
ResNet50 Adam 16 0.001 100
MobileNetV2 Adam 32 0.0001 50
Proposed Deep CNN RMSprop 16 0.0001 100
Moderate VGG16 Adam 32 0.0001 100
ResNet50 Adam 32 0.0001 100
MobileNetV2 Adam 32 0.0001 100
Proposed Deep CNN RMSprop 16 0.0001 100
Severe VGG16 Adam 32 0.0001 100
ResNet50 Adam 32 0.0001 100
MobileNetV2 Adam 32 0.0001 100
Proposed Deep CNN RMSprop 16 0.0001 100
 COMPUTER METHODS IN BIOMECHANICS AND BIOMEDICAL ENGINEERING: IMAGING & VISUALIZATION 7

tuning. When the dataset is small, this method can be overfitting and results in better accuracy. Out of total images,
adopted for medical image classification (Tajbakhsh et al. 70% are used for training and 30% are used for testing the
2017). Initial layers are frozen for all the models used and model. From the training set 10% of images are used for
the final fully connected layers are replaced with new fully validation. The number of images of normal and three classes
connected dense layer with 2-class classification task. The of glaucoma used for training and testing are shown in Table 4.
newly added layers are trained as a part of fine-tuning. The The deep CNN architecture model is implemented with
list of hyperparameters used for the VGG16, ResNet50 and Python programming language using the Keras library in
MobileNetV2 architectures is summarised in Table 3 below. Google Colab. The outputs from the layers of deep CNN are
visually represented to analyse the feature extracted from
within these layers. Since the most important features lie
Experimental results and discussion around the optic disc, the fundus image features are visually
represented to find out if the model is extracting the features
The deep CNN architecture proposed is trained and tested relevant to glaucoma or not. This helps to understand which
using the dataset mentioned in Table 1. We have a fundus features of the input image are detected and preserved in the
dataset of 295 normal images out of which we have taken feature maps obtained. While visualising from the first layer of
images in such a way that during training and validation the convolution to the output layer, the feature maps in the start­
class imbalance problem is eliminated. This helps to reduce ing layers detect finer details whereas feature maps closer to
the output layers detect global features. The first layer of acti­
Table 4. Total images used for training and testing the architecture. vation has 32 filters and hence 32 channel feature maps can be
Glaucoma classes Glaucoma images Normal images Total plotted. This number increases as the number of filters and
Mild 50 60 110 layers increases. Thus, we have plotted some feature maps in
Moderate 118 120 238 the first layer of convolution as depicted in Figure 6. The feature
Severe 137 140 277 maps convey that the convolution layers followed by max-

Figure 6. Visualization of the feature maps of first layer of activation. The first row is a feature map of mild glaucoma, the second row is moderate glaucoma, and the
third row is severe glaucoma. The feature maps are plotted for the 7th and 16th channels which are taken randomly out of the 32 channels in the first activation layer.
8 J. N K ET AL.

Figure 7. Accuracy curve and loss curve of mild glaucoma class.

Figure 8. Accuracy curve and loss curve of moderate glaucoma class.

Figure 9. Accuracy curve and loss curve of severe glaucoma class.

pooling layers are detecting the optic disc boundaries, brighter The training and validation accuracy curves and loss curves
optic cup region, blood vessels (though not relevant to glau­ of mild, moderate, and severe glaucoma classes are plotted as
coma), and retinal nerve fibre layers. The model can be trained shown in Figures 7, 8, 9 respectively as well as accuracy for all
in such a way that all the relevant features are extracted effi­ classes is calculated. For the mild glaucoma class, the training
ciently for a better discrimination of glaucoma from the normal accuracy obtained is 85.7% while validation accuracy is 87.8%.
class. Similarly, the training and validation accuracy for moderate and
 COMPUTER METHODS IN BIOMECHANICS AND BIOMEDICAL ENGINEERING: IMAGING & VISUALIZATION 9

Table 5. Training and validation accuracy of glaucoma dataset.

Glaucoma
class Training accuracy Validation accuracy Training loss Validation loss
Mild 85.71 87.88 0.31 0.36
Moderate 92.81 94.37 0.17 0.23
Severe 93.8 89.1 0.12 0.48

Figure 10. Confusion matrices for (a) mild glaucoma (b) moderate glaucoma (c) severe glaucoma.

Table 6. Results of performance metrics evaluated for the deep CNN architecture.
Glaucoma Precision
class Accuracy (%) Sensitivity (%) Specificity (%) (%) F1-score MCC
Mild 93.75 100 87.5 88.8 0.94 +0.88
Moderate 100 100 100 100 1 +1
Severe 93.75 100 83.3 90.9 0.95 +0.87

severe glaucoma classes are 92.8% and 94.3%, and 93.8% and depicted in Figure 10, accuracy, specificity, sensitivity, preci­
89.1%, respectively. These values are marked in Table 5. sion, F1 score and MCC have been calculated as shown in
Confusion matrices for all classes are computed to evalu­ Table 6. The values obtained help us to understand if the
ate the performance of the model. From these matrices model is discriminating between the glaucoma and normal
10 J. N K ET AL.

Table 7. Comparison of result of test accuracy for proposed deep CNN model and ImageNet trained architectures.
Glaucoma VGG16 ResNet50 MobileNetV2 Proposed deep CNN model
Class Test Accuracy (%) Test Accuracy (%) Test Accuracy (%) Test Accuracy (%)
Mild 84.09 81.82 75 93.75
Moderate 89.13 97.83 89.03 100
Severe 97.03 93.07 94.06 93.75

Table 8. Comparison of performance metrics evaluated for the existing models vs proposed model.
Model ACC (%) SE (%) SP (%) PRC (%) F1-score MCC
(Diaz-Pinto et al. 2019) 80.00 83.33 77.78 - - -
(Elangovan and Nath 2021) 96.64 96.07 97.39 97.74 - -
(Li et al. 2020) 85.20 84.80 85.50 - - -
(Alghamdi et al. 2017) 86.52 96.42 86.00 - - -
Proposed 93.75 100 87.5 88.8 0.94 +0.88

classes correctly. For a model to perform well, the range of glaucoma, the RNFL thickness (thinning of which is strongly
values for F1 score (which is the weighted average of sen­ correlated with the presence of glaucoma), can be prominently
sitivity or recall and precision) should be between 0 and 1, 0 seen in the moderate class but which appears obliterated in the
being the worst performance and 1 giving the best. It can severe class of glaucoma images.
be seen from Table 6 that mild, moderate and severe glau­
coma classes have values of 0.94, 1, and 0.95 respectively
indicating that the model is able to discriminate between Comparison with existing work
the classes well.
Further, related to another important metric, the Mathew’s Since a considerable body of the literature exists on fundus
Correlation Coefficient (MCC), if the binary classifier predicts image classification for glaucoma, the work reported here has
a majority of positive and negative instances accurately, it been compared with the existing literature. The following
generates a high score. The range of MCC is typically between Table 8 provides this comparison.
−1 and + 1, with former indicating the worst value and the It can be seen from the table that certain metrics computed
latter the best. Our model generated MCC of + 0.88, +1 and + using our model compare favourably with those of (Elangovan
0.87 for mild, moderate and severe glaucoma, respectively, and Nath 2021) who appear to have reported the best results.
which is consistent with other metrics. Overall, the accuracy, F1- However, they have not reported certain metrics such as F1-
score and MCC show reliable scores for predictions that accu­ score and MCC as done by us. It must also be emphasised that
rately classify the positive and negative classes. while these authors have used the publicly available datasets,
our dataset is a specialised one available with an eye institute.
Thus, a direct comparison of results may not be very
Comparison with ImageNet trained CNN architectures appropriate.

The mild, moderate, and severe classes of glaucoma are classi­

fied using state-of-the-art networks like VGG16, ResNet50, and
MobileNetV2 and compared with the proposed deep CNN Conclusion
model for all the classes. The input fundus images are resized Detection of glaucoma is important due to the progressive
to 224 × 224 * 3 before being input to all these architectures. nature of this ocular disease. If not detected and treated at
The visual representation of the outcome of the layers of deep the right time, glaucoma can lead to a loss of vision and later to
CNN model proposed gave us insight into the local features permanent blindness. In this paper, we implemented a deep
such as edges, corners of the optic disc (related to neuro-retinal CNN architecture with visualisation of features to distinguish
rim), and RNFL region within the fundus images. The perfor­ between glaucomatous and healthy fundus images. We
mance metrics are reported in Table 6. The ImageNet trained attempted visualisation between the layers of CNN to identify
CNN architectures used also were pretrained and finetuned to the features extracted and they turned out to be similar to
extract features and classify the fundus image datasets. The test those considered by the clinicians while diagnosing for glau­
accuracy obtained with these architectures for each class of coma. However in practice, the clinicians also perform other
glaucoma against the normal class and their comparison with tests such as HVF (Humphrey Visual Field) test to distinguish
the proposed CNN model are depicted in Table 7. The results between mild, moderate and severe glaucoma which are not
indicate that, in the mild class of glaucoma images the features built into the current model. Thus, there is a limit on the
are not so prominent leading to the model giving an accuracy accuracy of prediction which turned out to be still high at
of 93.75% which is better when compared to models such as 93.75% (on the lower side) in this work.
VGG16, ResNet50, and MobileNetV2. Further, while the moder­
ate class gives an accuracy of 100%, the severe class of glau­
coma images gave an accuracy of only 93.75% which is same as
Disclosure statement
that of the mild class. This can be understood from the images
shown in Figure 6 where a prominent feature relevant to No potential conflict of interest was reported by the author(s).
 COMPUTER METHODS IN BIOMECHANICS AND BIOMEDICAL ENGINEERING: IMAGING & VISUALIZATION 11

Notes on contributors Alghamdi HS, Tang HL, Waheeb SA, Peto T. 2017. Automatic optic disc
abnormality detection in fundus images: a deep learning approach.
Jisy N K is a Ph.D. student in the Department of Electrical and Electronics 17–24. 10.17077/omia.1042.
Engineering at Birla Institute of Technology and Science (BITS) Pilani, Anbalagan T, Nath MK, Vijayalakshmi D, Anbalagan A. 2023. Analysis of
Hyderabad Campus, India. She received her B.Tech degree in Electronics various techniques for ECG signal in healthcare, past, present, and
and Communication Engineering from AWH Engineering College, Kerala, future. Biomed Eng Adv. 6(April):100089. doi: 10.1016/j.bea.2023.
India in 2009 and her Masters degree in VLSI design from Anna University 100089.
Regional Centre, Coimbatore, India in 2013. She was honored with the Apostolopoulos ID, Apostolopoulos DJ, Papathanasiou ND. 2022. Deep
College Topper Award in 2010 and was a Rank holder in her ME VLSI learning methods to reveal important X-ray features in COVID-19 detec­
Design program in 2013. Her current research interests include biomedical tion: investigation of explainability and feature reproducibility. Reports.
image processing, application of machine learning, and deep learning for 5(2):20. doi: 10.3390/reports5020020.
fundus image processing. Bernardes R, Serranho P, Lobo C. 2011. Digital ocular fundus imaging: a
Dr. Md Hasnat Ali completed his Ph.D. from Birla Institute of Technology review. Ophthalmologica. 226(4):161–181. doi: 10.1159/000329597.
and Science (BITS) Pilani, Hyderabad Campus from the Department of Bock R, Meier J, Nyúl LG, Hornegger J, Michelson G. 2010. Glaucoma risk
Electrical and Electronics Engineering. He is currently the senior bio-statis­ index: Automated glaucoma detection from color fundus images. Med
tician at L V Prasad Eye Institute. He has been an integral part of the Image Anal. 14(3):471–481. doi: 10.1016/j.media.2009.12.006.
institute’s faculty since 2010, bringing his expertise and knowledge to Bogacsovics G, Toth J, Hajdu A, Harangi B. 2022. Enhancing CNNs through
advance research and statistical analysis in the field of ophthalmology. He the use of hand-crafted features in automated fundus image
has demonstrated exceptional proficiency in designing robust statistical classification. Biomed Signal Process Control. 76(October 2021):103685.
models, analyzing complex data sets, and interpreting results precisely. He doi: 10.1016/j.bspc.2022.103685.
received his MBA in Public Health Informatics from Jamia Hamdard, uni­ Bulut B, Kalin V, Gunes BB, Khazhin R. 2020. Deep learning approach for
versity, New Delhi. His research interests include circular and functional detection of retinal abnormalities based on color fundus images. In:
data analysis, computational statistics and machine learning, predictive Proceedings - 2020 Innovations in Intelligent Systems and Applications
modeling, and Big data analysis. He has published in over 150 international Conference, ASYU 2020. doi: 10.1109/ASYU50717.2020.9259870.
publications as an author or co-author. Coleman AL, Miglior S. 2008. Risk factors for glaucoma onset and
progression. Surv Ophthalmol. 53(6 SUPPL):3–10. doi: 10.1016/j.sur
Dr. Sirisha Senthil is a highly accomplished Ophthalmologist with over 22
vophthal.2008.08.006.
years of experience in the field of Glaucoma. She holds an MS in
Diaz-Pinto A, Morales S, Naranjo V, Köhler T, Mossi JM, Navea A. 2019. Cnns
Ophthalmology from Aravind Eye Hospital, Madurai, and an FRCS in
for automatic glaucoma assessment using fundus images: an extensive
Ophthalmology from Edinburgh. She is currently the Head of Glaucoma
validation. Biomed Eng Online. 18(1):1–19. doi: 10.1186/s12938-019-
service at L V Prasad Eye Institute, where she has been a faculty member
0649-y.
since 2007. Dr. Senthil specializes in managing refractory adult and pedia­
Dimitriou CD, Broadway DC. 2013. Pathophysiology of glaucoma.
tric glaucomas, with a particular interest in glaucoma drainage implants
Glaucoma: Basic Clin Perspect. 33–57. doi: 10.2217/EBO.12.421.
and secondary glaucomas and genetics of inherited childhood glaucoma.
Dong K, Zhou C, Ruan Y, Li Y. 2020. MobileNetV2 model for image
She has published over 186 scientific papers in peer-reviewed journals,
classification. In: Proceedings - 2020 2nd International Conference on
authored several book chapters and received multiple awards in national
Information Technology and Computer Application, ITCA 2020 p.
and international fora for her work, including Dr. Vengal Rao Medal in 2012,
476–480. 10.1109/ITCA52113.2020.00106.
IJO Silver award in 2016, and the American Academy of Ophthalmology
E H, II PR, DR A. 1993. Clinical decisions in glaucoma. St Louis: The CV
Achievement award in 2018.
Mosby Co.
M.B. Srinivas received his Ph.D. in Electrical Engineering from the Indian Elangovan P, Nath MK. 2021. Glaucoma assessment from color fundus
Institute of Science, Bangalore, India. He is currently a Professor in the images using convolutional neural network. Int J Imaging Syst
Department of Electrical and Electronics Engineering at Birla Institute of Technol. 31(2):955–971. doi: 10.1002/ima.22494.
Technology and Science (BITS) Pilani. His research interests include Elangovan P, Nath MK. 2022. En-ConvNet: a novel approach for glaucoma
Computer Arithmetic and Architectures, Medical Device Design and Deep detection from color fundus images using ensemble of deep convolu­
Learning Models for Medical Image Processing. He was a recipient of tional neural networks. Int J Imaging Syst Technol. 32(6):2034–2048. doi:
Microsoft Research ’Digital Inclusion’ award in 2006 and Stanford 10.1002/ima.22761.
Medicine ’MedTech Innovation’ award in 2016. Elangovan P, Nath MK, Mishra M. 2020. Statistical parameters for glaucoma
detection from color fundus images. Procedia Comput Sci. 171
(2019):2675–2683. doi: 10.1016/j.procs.2020.04.290.
Hussain SA, Holambe AN. 2015. Automated detection and classification of
ORCID glaucoma from eye fundus images: a survey. Int J Comput Sci Inf
Technol. 6(2):1217–1224.
Jisy N K http://orcid.org/0000-0002-3788-719X
Jisy NK, Md HA, Senthil S, Srinivas MB. 2024. Localization of region of
Sirisha Senthil http://orcid.org/0000-0003-1748-8850
interest from retinal fundus image for early detection of glaucoma.
Located: Proc Ninth Int Congr Inf Commun Technol: ICICT 2024. 6.
London, Singapore: Springer.
References Johora FT, Mahbub -Or-Rashid M, Yousuf MA, Saha TR, Ahmed B. 2020.
Diabetic retinopathy detection using PCA-SIFT and weighted decision
Abbas Q. 2017. Glaucoma-Deep: detection of Glaucoma eye disease on tree. January. 2020:25–37. doi: 10.1007/978-981-13-7564-4_3.
retinal fundus images using Deep learning. Int J Advan Comput Sci Appl. Juneja M, Singh S, Agarwal N, Bali S, Gupta S, Thakur N, Jindal P. 2020.
8(6):41–45. doi: 10.14569/ijacsa.2017.080606. Automated detection of Glaucoma using deep learning convolution
Alam Miah MB, Yousuf MA. 2018. Analysis the significant risk factors on type network (G-net). Multimedia Tools Appl. 79(21–22):15531–15553. doi:
2 diabetes perspective of Bangladesh. Diabetes Metab Syndr. 12 10.1007/s11042-019-7460-4.
(6):897–902. doi: 10.1016/j.dsx.2018.05.012. Kamal MS, Dey N, Chowdhury L, Hasan SI, Santosh KC. 2022. Explainable AI
Alayón S, Hernández J, Fumero FJ, Sigut JF, Díaz-Alemán T. 2023. for glaucoma prediction analysis to understand risk factors in treatment
Comparison of the performance of convolutional neural networks planning. IEEE Trans Instrum Meas. 71:1–9. doi: 10.1109/TIM.2022.
and vision transformer-based systems for automated glaucoma 3171613.
detection with eye fundus images. Appl Sci. 13(23):12722. doi: 10. Khan KB, Khaliq AA, Jalil A, Iftikhar MA, Ullah N, Aziz MW, Ullah K, Shahid M.
3390/app132312722 . 2019. A review of retinal blood vessels extraction techniques: challenges,
12 J. N K ET AL.

taxonomy, and future trends. Pattern Anal Applic. 22(Issue 3):767–802. Rao HBL, Sekhar GC, Babu GJ, Parikh RS. 2009. Clinical measurement and
Springer London. doi: 10.1007/s10044-018-0754-8. categorization of optic disc in glaucoma patients. Indian J Ophthalmol.
Khan A, Sohail A, Zahoora U, Qureshi AS. 2020. A survey of the recent 57(5):361–364. doi: 10.4103/0301-4738.55075.
architectures of deep convolutional neural networks. Artif Intell Rev. 53 Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang Z,
(Issue 8):5455–5516. Springer Netherlands. doi: 10.1007/s10462-020- Karpathy A, Khosla A, Bernstein M, et al. 2015. ImageNet large scale
09825-6. visual recognition challenge. Int J Comput Vision. 115(3):211–252. doi:
Leandrou S, Lamnisos D, Bougias H, Stogiannos N, Georgiadou E, 10.1007/s11263-015-0816-y.
Achilleos KG, Pattichis CS. 2023. A cross-sectional study of explainable Saiteja KV, Kumar PM, Chandra PS, Jisy NK, Ali MH, Srinivas MB. 2021.
machine learning in Alzheimer’s disease: diagnostic classification using Glaucoma - automating the cup-to-disc ratio estimation in fundus
MR radiomic features. Front Aging Neurosci. 15(June):1–11. doi: 10.3389/ images by combining random walk algorithm with otsu thresholding.
fnagi.2023.1149871. Int IEEE/EMBS Conf Neural Eng, NER. 2021(May):154–157. doi: 10.1109/
Lecun Y, Bengio Y, Hinton G. 2015. Deep learning. Nature. 521 NER49283.2021.9441435.
(7553):436–444. doi: 10.1038/nature14539. Sandler M, Howard A, Zhu M, Zhmoginov A, Chen LC. 2018. MobileNetV2:
Li L, Xu M, Liu H, Li Y, Wang X, Jiang L, Wang Z, Fan X, Wang N. 2020. A inverted residuals and linear bottlenecks. In: Proceedings of the IEEE
large-scale database and a CNN model for attention-based glaucoma Computer Society Conference on Computer Vision and Pattern
detection. IEEE Trans Med Imaging. 39(2):413–424. doi: 10.1109/TMI. Recognition p. 4510–4520. 10.1109/CVPR.2018.00474.
2019.2927226. Sarhan MH, Nasseri MA, Zapp D, Maier M, Lohmann CP, Navab N, Eslami A.
2020. Machine learning techniques for ophthalmic data processing: a
McMonnies CW. 2017. Historial de glaucoma y factores de riesgo.
review. IEEE J Biomed Health Inf. 24(12):3338–3350. doi: 10.1109/JBHI.
[Glaucoma history and risk factors]. J Optom. 10(2):71–78. Spanish. 10.
2020.3012134.
1016/j.optom.2016.02.003.
Simonyan K, Zisserman A. 2015. Very deep convolutional networks for
Militello C, Prinzi F, Sollami G, Rundo L, La Grutta L, Vitabile S. 2023. CT
large-scale image recognition. In: 3rd International Conference on
radiomic features and clinical biomarkers for predicting coronary artery
Learning Representations, ICLR 2015 - Conference Track Proceedings,
disease. Cognit Comput. 15(1):238–253. doi: 10.1007/s12559-023-
San Diego, CA, USA, p. 1–14.
10118-7 .
Sudhan MB, Sinthuja M, Pravinth Raja S, Amutharaj J, Charlyn Pushpa
Nandhini Abirami R, Durai Raj Vincent PM, Srinivasan K, Tariq U, Chang CY,
Latha G, Sheeba Rachel S, Anitha T, Rajendran T, Waji YA, Abdulhay E.
Sarfraz DS. 2021. Deep CNN and deep GAN in computational visual 2022. Segmentation and classification of glaucoma using U-Net with
perception-driven image analysis. Complexity. 2021:1–30. doi: 10.1155/ deep learning model. J Healthc Eng. 2022:1–10. doi: 10.1155/2022/
2021/5541134. 1601354.
Oh S, Park Y, Cho KJ, Kim SJ. 2021. Explainable machine learning model for Sun Q, Lin X, Zhao Y, Li L, Yan K, Liang D, Sun D, Li ZC. 2020. Deep learning
glaucoma diagnosis and its interpretation. Diagnostics. 11(3):1–14. doi: vs. Radiomics for predicting axillary lymph node metastasis of breast
10.3390/diagnostics11030510. cancer using ultrasound images: don’t forget the peritumoral region.
Orlando JI, Prokofyeva E, Del Fresno M, Blaschko MB. 2017. Convolutional Front Oncol. 10(January):1–12. doi: 10.3389/fonc.2020.00053.
neural network transfer for automated glaucoma identification. 12th Tajbakhsh N, Shin JY, Gurudu SR, Hurst RT, Kendall CB, Gotway MB, Liang J.
Int Symp Med Inf Proc Anal. 10160:101600U. doi: 10.1117/12.2255740. 2017. Convolutional neural networks for medical image analysis: full
Park K, Kim J, Lee J. 2020. Automatic optic nerve head localization and training or fine tuning? ArXiv. 35(5):1299–1312. doi: 10.1109/TMI.2016.
cup-to-disc ratio detection using state-of-the-art deep-learning 2535302.
architectures. Sci Rep. 10(1):3–12. doi: 10.1038/s41598-020-62022-x. Wei P. 2021. Radiomics, deep learning and early diagnosis in oncology.
Prinzi F, Militello C, Scichilone N, Gaglio S, Vitabile S. 2023. Explainable Emerging Top Life Sci. 5(6):829–835. doi: 10.1042/ETLS20210218.
machine-learning models for COVID-19 prognosis prediction using clin­ Yosinski J, Clune J, Bengio Y, Lipson H. 2014. How transferable are features
ical, laboratory and radiomic features. IEEE Access. 11 in deep neural networks? Adv Neural Inf Process Syst. 4
(October):121492–121510. doi: 10.1109/ACCESS.2023.3327808. (January):3320–3328.