OCT (Optical Coherence Tomography) Image

for Disease Detection Using Transfer Learning

Techniques.
Department of Electronics and Communication Engineering, NIT Jalandhar, Punjab, (INDIA)

Deepti Kakkar Srajal Dwivedi Sakirti

NIT Jalandhar NIT Jalandhar NIT Jalandhar
[email protected] [email protected] [email protected]
Gitanshu Lucky Verma
NIT Jalandhar NIT Jalandhar
[email protected] [email protected]
Abstract: Investigating state-of-art diagnostic To achieve early detection, Optical Coherence
techniques, this study delves into advanced Tomography (OCT) serves as a non-invasive
methods for eye disease identification. For this medical technique for studying tissue structures.
Optical Coherence Tomography stands as a
pivotal implementation for the comprehensive OCT uses low-coherent light and interferometry to
transverse imaging of retinal tissues. In the initial acquire axial (A-scan) and cross-sectional (B-scan)
stages of OCT image classification for retinal images, demanding complex software techniques for
diseases, traditional non-deep learning noise reduction, segmentation, and classification.
approaches were employed. However, these A-scan determines depth by measuring the echo
methods exhibited susceptibility to noise, time delay along a single line, but B-scan creates a
compromising their overall effectiveness. detailed cross-sectional image by integrating
Recognizing the limitations, a subsequent shift numerous A-scan results. These complimentary
towards deep learning approaches was scanning approaches enable OCT to provide high-
undertaken, showcasing promising results. resolution, 3-D images, making it a useful tool for
Nevertheless, inherent drawbacks persisted non-invasive imaging in a variety of medical
within this methodology. To address these domains, particularly in ophthalmology for
challenges and enhance the classification comprehensive analysis of retinal structures.
performance for retinal diseases, explicitly AMD We collected about 95 papers from IEEE Explore,
(age-related), CNV, DME, and Normal cases, Scopus, BioMed, PubMed, and Google Scholar.
fig.[I] this paper proposes the integration of a After an extensive search approximately, 17 studies
transfer learning approach fig. [II]. This were identified that met the criteria of our study. The
innovative strategy aims to leverage pre-existing retina is a fragile and important layer at the back side
knowledge from a source domain to improve the of the eye that captures and processes light signals.
classification accuracy and robustness of the However, retinal abnormalities affecting different
model for OCT image analysis. parts of the retina can range from moderate vision
Keywords: Optical coherence tomography, problems to blindness. Early detection and
Classification, Deep learning, CNN, Transfer intervention by eye specialists are crucial in
learning, medical imaging. reducing the impact of these illnesses.

A. CLASSIFICATION OF RETINAL
DISEASES
I. INTRODUCTION
Retinal Diseases cause damage to retina of human
Detecting early-stage retinal diseases poses a being.
significant challenge, as these conditions advance
over time. Morphological alterations in the retina  Age-related Macular Degeneration (AMD)
often precede noticeable symptoms, making it  Diabetic Macular Edema (DME)
imperative to identify subtle changes as individuals  Choroidal Neovascularization Disease
age. (CNV)
 Normal Retina

1|Page
 Fig I: Diseases under consideration

 Age-Related Macular Degeneration

(AMD) manifests as a condition
predominantly affecting the central visual
field, particularly among individuals aged
50 and beyond. While it doesn't precipitate
complete visual impairment, it
significantly impedes rudimentary Fig II: Process for retinal disease detection
activities such as reading. If left
unaddressed, AMD can evolve into either
dry-aged macular or wet-aged macular
B. APPROACH
pathologies.
 Diabetic Macular Edema arises because of Deep learning approaches for image classification
diabetic conditions, resulting in the accrual have shown considerable promise. In contrast,
of fluid within the macula. When the traditional machine learning methods involve the
macula undergoes swelling due to fluid manual extraction of features, which are then passed
accumulation, it is denoted as Edema. to various classifiers. This approach is notably
 Choroidal Neovascularization denotes the susceptible to noise and tends to be less efficient.
aberrant proliferation of new vascular
networks in the choroid, situated posterior In our current project, we adopt a transfer learning
to the retina. Symptoms of choroidal approach, leveraging a pre-trained model to address
neovascularization encompass distorted a novel problem. Transfer learning is a sophisticated
vision, the manifestation of undulating machine learning method that repurposes a model
patterns in central vision, and the initially designed for one task to excel at a related
emergence of a grey or black void within yet distinct task.
the central visual field. This condition is The rationale behind choosing transfer learning
frequently linked to age-related macular stems from the limitations of conventional deep
degeneration. learning methods employed by other researchers.
 In the scenario of an unimpaired retina, we These methods often demand a substantial amount
would be privy to data derived from of optical coherence tomography data for effective
visually robust or unaffected ocular organ. model training. To circumvent this challenge, we
embrace transfer learning, which not only requires
comparatively less data but also accelerates the
training process, thereby enhancing overall model
performance.

By leveraging transfer learning, we optimize

efficiency, reduce data requirements, and expedite
the model training process, leading to improved
performance outcomes.

2|Page
 II. Related Research Work network trained as a B-scan feature extractor is one
of the two 2-D image machine learning
The conventional problem in categorizing medical classification techniques used by the architecture to
images is the limitation of data available and the data create a 2-D feature map of an OCT volume. For 2-
imbalance. The over-availability of healthy ones, D feature maps, the volume-level classifiers are
above the class in the training data, influences CNN with/without an attention mechanism and
results. To resolve this problem, the techniques used support vector machines. The duke dataset, which is
are the transfer learning approach, augmentation of accessible to the public, and a private clinical OCT
given data, and synthesis of data. Another drawback volume dataset are used to validate the suggested
of deep learning is the incapability to explain the approaches. The results of the studies demonstrate
reasoning of obtained results. Recent work has been that the suggested methods beat the most advanced
done in the classification of OCT Images for disease techniques and have the potential to categorize OCT
detection along with approaches/methods, dataset, volumes efficiently and accurately [16].
parameters used for evaluation of the performance
of the method are also shown in this literature
review.
In the paper [11], a novel Random Forest algorithm-
ReLayNet is used to create the retinal layer based method for classifying eye diseases is
segmentation maps for the proposed Layer Guided proposed. The suggested approach makes use of
Convolutional Neural Network (LGCNN) approach manually created, automatically quantified
[15]. Two lesion-related retinal layers are then clinically driven aspects of AMD, DME, and healthy
retrieved from these maps. The lesion-related data individuals. In both datasets, the Random Forest
from the lesion-related layers is then used to create approach outperforms other machine learning
two subnetworks. To merge the last connected layers algorithms in terms of accuracy. For the
of two subnetworks for the final classification of classification approach, the AUC value is likewise
retinal diseases, a fully connected layer is used at the very high (0.99) with a tiny standard deviation
end. With gains in sensitivity over the VGG16 model (0.001). The findings demonstrate a strong
of 4.19%, 44.87%, 11.17%, and 11.14% in four correlation between the thicknesses and pathologies
categories (CNV, drusen, DME, and normal retina), of the layers and the projected diseases. Compared
respectively, the LGCNN technique [15] performs to state-of-the-art procedures, the suggested method
best in most of the evaluation criteria. The suggested [11]is more accurate, although it may have certain
approach gives ophthalmologists a more drawbacks that require more investigation. The
understandable and straightforward diagnostic authors intend to collaborate with ophthalmologists
foundation. to create data sets with severity levels so that they
can be predicted in subsequent research.
In an OCT image dataset,[4] very few B-scans
shows the symptoms/ manifestation due to diseases In medical fields, imbalanced datasets, particularly
at an intermediate level. weights are assigned on in OCT images, pose a challenge for machine
discriminative information level. Compared to 3D learning models. Limited data on diseases like
voxel-based CNN classification, the B-scan cancer or brain tumours compared to healthy cases
attentive method showed better performance as the affects model performance. Researchers propose
clinical data do not lose during processing. B scan transfer learning and Generative Adversarial
attentive method showed an improvement of 8% for Networks (GAN) to address this issue. GANs
kappa value,3.5% for accuracy etc over traditional generate artificial data to balance or augment
methods. B-scan attentive approach can also be datasets. Most research focuses on segmenting
extended for severity grading of OCT images. [4] retinal layers, aiding in disease detection, severity
grading, and progression study. AMD is a priority
To automatically identify macular disorders, this eye disease, and early detection is crucial to prevent
paper [16] presents work on an auto-encoder system vision loss. Automatic classification of OCT scans,
for classifying OCT volumes. The B-scan feature especially for diseases like AMD and DME, is vital.
extractor, 2-D feature map generator, and volume- This often involves binary classification (e.g.,
level classifier are the three components that make normal vs. AMD) to enhance diagnosis efficiency.
up the framework. A convoluted neural

In medical fields, imbalanced datasets, particularly cancer or brain tumours compared to healthy cases
in OCT images, pose a challenge for machine affects model performance. Researchers propose
learning models. Limited data on diseases like transfer learning and Generative Adversarial

3|Page
Networks (GAN) to address this issue. GANs eye disease, and early detection is crucial to prevent
generate artificial data to balance or augment vision loss. Automatic classification of OCT scans,
datasets. Most research focuses on segmenting especially for diseases like AMD and DME, is vital.
retinal layers, aiding in disease detection, severity This often involves binary classification (e.g.,
grading, and progression study. AMD is a priority normal vs. AMD) to enhance diagnosis efficiency.

Table I: RECENT WORK ON OCT IMAGE CLASSIFICATION FOR DISEASE DETECTION

Author/year &
Method/Approach Datasets Accuracy
publication

Huang et al. (2018), Layer guided (LG-CNN) 108312 images (CNV: 37,206, Accuracy: 88% - 89%
CELL [15] for classification of CNV, DME:11,349, AMD
DME, and AMD diseases. Drusen:8,617, Normal:51,140).

Das et al., (2020), IEEE B-scan attentive CNN, Dataset by DUIA volume of 269 Accuracy - 93.2
Signal Processing Letters Classifying Normal, AMD, AMD and 115 control subjects
[4] DRUSEN and DME. having 100 B-scans per volume.
Another Dataset of 148 images
by NEH containing 50 DME
and normal, 48 dry AMD
volumes, with 19, 25, 31, and
61 B-scans per volume.
P. Seebock et al, IEEE deep auto-encoder with Public & Private Datasets. Accuracy- 81.40
Trans. Med. Imaging [16] support vector machine
(SVM). For Early and Late
AMD.
Sarki, R., Ahmed, K., Used transfer learning and Pre-processing is done only for Accuracy - 91.43
Wang, (2021) Image Pre- CNN Models for fundus images and don’t have
processing.[7] classification. any data regarding OCT images.

Miere, A., Capuano, V., The classification is done Fundus autofluorescence Accuracy - 92
(2021). Deep learning- based on two classes i.e., images, while effective for
based classification using genetic or age related specific classification, may not
fundus autofluorescence macular (AMD). be optimal for identifying other
imaging.[8] diseases.
Wang, J., Yang, L., Classification can be done This paper focuses solely on the Accuracy - 94
(2020). Multi-label by performing canny edge classification of two classes of
classification of fundus detection followed by retinal diseases, without
images with Efficient- feature extraction. extending its scope to
Net. IEEE, 212499- encompass the broader spectrum
212508.[9] of retinal conditions.
Hussain, M. A., Bhuiyan, This paper employs Graph In this paper they have used Accuracy - 89.9
A. Classification of Convolutional Neural coloured images for
healthy and diseased Network to assess the classification of diseases which
retina using SD-OCT severity of retinopathy. makes it difficult to identify
imaging and Random disease accurately.
Forest algorithm,
e0198281.[10]
Ai, Z., Huang, X., Feng, This paper introduces a Accuracy obtained is less by Accuracy - 84
J., (2022). FN-OCT: Surrogate-Assisted Model implementing this model.
Disease detection based on CNN to detect
algorithm for retinal OCT diseases related to the
based on a fusion retina.
network. Used in
Neuroinformatic, 16.[11]

4|Page
III. OVERVIEW OF TRANSFER involves a fully connected layer integrated with a
LEARNING APPROACH classifier.

The transfer learning-based approach for image In the realm of medical image analysis, diverse
classification has yielded highly promising and models are recommended to enhance disease
encouraging results. During the acquisition of detection and mitigate challenges such as limited
Optical Coherence Tomography (OCT) images, training samples, overlapping classes, overfitting,
motion artifacts and multiple scattering introduce underfitting, high-dimensional feature sets, and
speckle noise, imparting a granular appearance to imbalanced datasets.
the images. The presence of speckle noise not only
complicates clinical analysis for medical experts but The evaluation of a classification model entails a
also degrades the performance of computer-aided confusion matrix, encompassing metrics such as true
diagnosis (CAD) systems. Consequently, it becomes positives, true negatives, false positives, and false
imperative to mitigate speckle noise during data pre- negatives. The ramifications of detecting a disease
processing. when none is present or missing a disease when it
exists are substantial for patients. In the case of true
In situations where data is limited, the subsequent positives, the model accurately identifies the
step post pre-processing involves data presence of a disease. Conversely, for true negatives,
augmentation. The meticulously prepared dataset the model correctly identifies the absence of a
then serves as the input for training the bespoke disease. False positives occur when the model
transfer learning model. Within each layer of the erroneously detects a disease in the absence of one,
Convolutional Neural Network (CNN) model, while false negatives transpire when the model fails
convolution operations are executed with kernels of to detect a disease that is present.
varying sizes. The number of kernels or filters in
each layer is also deliberately varied. Over the Here, mentioned in table II, we have implemented
course of multiple convolutional layers, the model three models VGG-16, VGG-19 and InceptionV3
assimilates low-level and high-level features and have obtained astonishing results (fig [III] and
intrinsic to the dataset. The concluding stage fig [IV]) on the classification of different diseases.

Table II: RESULTS USING DIFFERENT DEEP LEARNING MODELS

Method Keras-CNN Diseases Precision Recall F1-Score Support

Used/Model Accuracy

VGG-16 Accuracy – Normal 0.97 1.00 0.99 39

94.230 CNV 0.90 0.95 0.92 39
Optimizer - DME 0.92 0.87 0.89 39
Adam DRUSEN 0.97 0.95 0.96 39

InceptionV3 Accuracy – Normal 0.89 1.00 0.94 39

91.025 CNV 0.90 0.92 0.91 39
Optimizer - DME 0.89 0.82 0.85 39
RMSprop DRUSEN 0.97 0.90 0.93 39

VGG-19 Accuracy – Normal 0.81 0.98 0.89 44

82.386 CNV 0.78 0.91 0.84 44
Optimizer - DME 0.88 0.68 0.77 44
Adam DRUSEN 0.84 0.73 0.78 44

5|Page
 Fig III: Result using VGG-19

Fig IV: Result using VGG-16

IV. CONCLUSION for the model. To solve this problem, we have

applied the transfer learning model, which is an
Nowadays, people are exposed to screens for long advanced technique in deep learning. Using transfer
hours. This leads to a rise in retinal diseases. learning, fully convoluted neural networks can
Therefore, it is necessary to detect retinal diseases leverage pre-trained networks to train faster and
early. To do this, we need a practical CAD perform better when the data is scarce. We can also
(computer-Aided diagnostic) system that can help
create synthetic data with the help of GAN, a
ophthalmologists diagnose the disease.
cutting-edge technique. Various models are
A few years ago, researchers applied a non-deep suggested in medical image analysis to improve
learning method for classifying OCT Images to disease detection and deal with problems like
detect Retinal diseases. This method involved insufficient training data, class imbalance, and
manual feature extraction, which made it very prone model overfitting and underfitting.
to noise. This is the reason why the deep learning
method is preferred, as it produces very positive and
V. REFERENCES
hopeful outcomes. In this method, we remove the
speckle noise in pre-processing and achieve the [1] K. Patil, A. Mahajan, S. Balamurugan, P.
desired result. Arulmozhivarman and R. Makkar, "Development of
Signal Processing Algorithm for Optical Coherence
One of the main challenges of deep learning
approach is that it needs a lot of OCT training data

6|Page
Tomography", Proc. 2020 IEEE Int. Conf. Commun. [10] Hussain, M. A., Bhuiyan, A., D. Luu, C.,
Signal Process. ICCSP 2020, pp. 1283-1287, 2020. Theodore Smith, R., H. Guymer, R., Ishikawa, H., ...
& Ramamohan Arao, K. (2018). Classification of
[2] P Nagaraj, V Muneeswaran, B Sunethra, C healthy and diseased retina using SD-OCT imaging
Sreeya, U Dhannushree, “A Comparative Analysis and Random Forest algorithm. PloS one, 13(6),
of Retinal Disease Image Classification for OCT e0198281.
Using Deep Learning Techniques”, Publisher 2023
IEEE. [11] Ai, Z., Huang, X., Feng, J., Wang, H., Tao, Y.,
Zeng, F., & Lu, Y. (2022). FN-OCT: Disease
[3] Jahida Subhedar, Anurag Mahajan, “A Review detection algorithm for retinal optical coherence
on Recent Work on OCT Image Classification for tomography based on a fusion network. Frontiers in
Disease Detection” SIU Pune India. Published in Neuroinformatic, 16.
2022 OPJU International Technology Conference
on Emerging Technologies for Sustainable [12] Sharan, E. S., Kumar, K. S., & Madhuri, G.
Development (OTCON). (2021, July). Conceal Face Mask Recognition Using
Convolutional Neural Networks. In 2021 6th
[4] V. Das, E. Prabhakararao, S. Dandapat, and P. International Conference on Communication and
K. Bora, “B-Scan Attentive CNN for the Electronics Systems (ICCES) (pp. 1787-1793).
Classification of Retinal Optical Coherence IEEE.
Tomography Volumes,” IEEE Signal Process. Lett.,
vol. 27, pp. 1025–1029, 2020, doi: [13] Das, A., Giri, R., Chourasia, G., & Bala, A. A.
10.1109/LSP.2020.3000933. (2019, July). Classification of retinal diseases using
transfer learning approach. In 2019 International
[5] D. S. Kermany et al., “Identifying Medical Conference on Communication and Electronics
Diagnoses and Treatable Diseases by Image-Based Systems (ICCES) (pp. 2080-2084). Publisher: IEEE.
Deep Learning,” Cell, vol. 172, no. 5, pp. 1122-
1131.e9, 2018, doi: 10.1016/j.cell.2018.02.010. [14] Mehmood Nawaz, Adilet Uvaliyev, Khadija
Bibi, Hao Wei, Sai Mu Dalike Abaxi, Anum Masood,
[6] M. A. Hussain et al., “Classification of healthy Peilun Shi, Ho-Pui Ho, Wu Yuan. “Unravelling the
and diseased retina using SD-OCT imaging and complexity of Optical Coherence Tomography
Random Forest algorithm,” PLoS One, vol. 13, no. image segmentation using machine and deep
6, pp. 1–17, 2018, doi: learning techniques”.
10.1371/journal.pone.0198281.
[15] L. Huang et al., “Automatic Classification of
[7] Sarki, R., Ahmed, K., Wang, H., Zhang, Y., Ma, Retinal Optical Coherence Tomography Images with
J., & Wang, K. (2021). Image preprocessing in Layer Guided Convolutional Neural Network,”
classification and identification of diabetic eye IEEE Signal Process. Lett., vol. 26, no. 7, pp. 1026–
diseases. Data Science and Engineering, 6(4), 455- 1030, 2019, Doi: 10.1109/LSP.2019.2917779.
471.
[16] P. Seebock et al., “Unsupervised Identification
[8] Miere, A., Capuano, V., Kessler, A., Zambrowski, of Disease Marker Candidates in Retinal OCT
O., Jung, C., Colantuono, D., ... & Souied, E. (2021). Imaging Data,” IEEE Trans. Med. Imaging, vol. 38,
Deep learning-based classification of retinal no. 4, pp. 1037–1047, 2019, Doi:
atrophy using fundus autofluorescence imaging. 10.1109/TMI.2018.2877080.
Computers in Biology and Medicine, 130, 104198.
[17] T. Tsuji et al., “Classification of optical
[9] Wang, J., Yang, L., Huo, Z., He, W., & Luo, J. coherence tomography images using a capsule
(2020). Multi-label classification of fundus images network,” BMC Ophthalmol, vol. 20, no. 1, pp. 1–9,
with efficient net. IEEE Access, 8, 212499-212508. 2020, Doi: 10.1186/s12886-020-01382-4.

7|Page