This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

Super-resolution of medical images using real

ESRGAN
P. Nandal1, Sudesh Pahal2, Ashish Khanna3, Placido Rogério Pinheiro4
1
Department of Computer Science and Engineering, Maharaja Surajmal Institute of Technology, New Delhi 110058, India
2
Department of Electronics and Communication Engineering, Maharaja Surajmal Institute of Technology, New Delhi 110058, India
3
Department of Computer Science and Engineering, Maharaja Agrasen Institute of Technology, New Delhi 110086, India
4
Graduate Program in Applied Informatics, University of Fortaleza, Ceará 60020-181, Brazil

Corresponding author: Sudesh Pahal (e-mail: sudeshpahal@msit.in).

ABSTRACT Rich details in an image are constantly vital for medical image analysis to detect a broad extent
of medical ailments. The diagnosis will be best served if the image is accessible in high resolution and the
small details are preserved. Image super-resolution techniques based on deep learning can assist us in
extracting spatial features from a low-resolution image captured with current technologies. The updated
variant of the super-resolution technique known as Real Enhanced Super-Resolution Generative Adversarial
Networks (Real-ESRGAN), which produces 2D real-world images with great perceptual quality, is used in
the present work. We investigate the suggested approach using four distinct medical image types: (1) brain
MRI images from the BraTS dataset; (2) dermoscopy images from the ISIC skin cancer dataset; (3) cardiac
ultrasound images from the CAMUS dataset; and (4) chest x-rays images from the MIMIC-CXR dataset. The
employed architecture achieves improved visual results in comparison to the alternative innovative
techniques for super-resolution. The observed findings are evaluated and contrasted both qualitatively and
quantitatively with conventional approaches in terms of PSNR, SSIM, and MSE, and an improvement of up
to 12% is obtained.

INDEX TERMS Deep learning, Generative adversarial network, Medical image enhancement, Real-
ESRGAN.

I. INTRODUCTION acquisition times, physiological movement, and the

Super-resolution (SR) is a difficult approach. A low- patient's acceptable radiation dose. The upgradation of
resolution (LR) image is reconstructed in high-resolution image-acquiring equipment is not only costly, but it is also
(HR) using SR techniques. Compared to LR images, HR limited in its ability to increase the quality of images [3].
images have more amplified information structures. Image Therefore, medical images can be post-processed and
SR, in technical terms, is a method of reconstructing the reconstructed in HR using SR techniques [4] without the
high-frequency components of LR images. SR techniques need for such an expensive image acquisition approach.
have been put out in recent years for both natural and Upscaling medical images is a challenge mainly because
medical images. Medical imaging serves as a vital tool for most of them (CT, MRI, and X-ray) have relatively poor
medical professionals to diagnose and treat patients. For contrast. Secondly, since SR is an ill-posed inverse
quantitative analysis and accurate diagnosis, isotropic problem, there are a number of SR outputs for every LR
medical images in HR having detailed textural and input [5]. Thirdly, medical image SR does not provide
structural information are required. Small anatomical ground truth training datasets like natural image SR does.
structures and other minute information in medical imaging The primary goal of medical image enhancement efforts
might reveal vital information that can be helpful for has been to enhance the ability to perceive blurry or unclear
diagnosis. For instance, in brain MRI, the minute structural images visually. The discontinuities of image intensity
characteristics surrounding a tumor aid in the diagnosis of curves are represented by edges. A strong edge
the tumor's origin and rate of growth [1, 2]. Similar to this, enhancement method is therefore necessary to process
accurately identifying tiny arteries in retinal images aids in these discontinuities in an image.
the diagnosis of inflammation of nerves, a sign of Over the years, a great deal of research has been done on
hypertensive retinopathy. Obtaining such images in clinical deep learning-based SR models. Lately, deep learning
practice is difficult as several factors affect the quality of approaches for SR of images have garnered more attention
medical images, including hardware specifications, [6-10]. Generally, these models show breakthroughs in

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

performance on different SR benchmarks. SR The layout of the remaining paper is, Section 2 illustrates
reconstruction algorithms can be categorized into three the related work existing in the literature. In Section 3, the
divisions, namely, Convolutional Neural Network (CNN)- proposed work is presented. The constitution of Section 4
based methodology (e.g., SRCNN (Super Resolution is the results analysis and Section 5 is discussion. Finally,
Convolutional Neural Network) [11], FSRCNN [12] and Section 6 concludes the work of this paper as well as the
ESPCN [13]), ResNet-based methodology (e.g., VDSR future development directions.
methodology [14] and EDSR methodology [15]), and
GAN-based methodology (e.g., SRGAN [16, 17] and II. RELATED WORK
DoubleGAN [18]). Fig. 1 displays the classification Deep Neural Networks (DNN) have shown exemplary
approaches of SR based on deep learning present in the results in image analysis and processing. These strategies
literature from early techniques employing CNNs (e.g., have performed better than many conventional techniques
SRCNN [11, 19]) to the current SR methods utilizing [24]. Deep learning was introduced to the area of image SR
transformers (e.g., TTSR [20]). by Dong et al. [25]. The authors recommended the SRCNN
model, which learned the mapping relationship amid LR
and HR images using a three-layer CNN. Features from LR
images were extracted using a feed-forward CNN. Bilinear
interpolation was then used to up-sample the features.
SRCNN outperformed the conventional techniques in terms
of reported results. Since SRCNN was the initial model to
introduce deep learning for image SR, it has served as a
fundamental to the advancement of image SR and evolved
into a groundbreaking study of deep learning-based SR
techniques. Deep recursive CNN was utilized by Kim et al.
[26] to extract complex features for the purpose of image
SR. The three components of the approach they used are as
FIGURE 1. Classical methods of SR algorithms on the basis of deep follows: basic features were extracted using the embedded
learning.
network, deep features were extracted using the inference
network, and features were mapped back to the HR image
In order to recapture edge and minute texture information
by the reconstruction layer. The deep recursive method is
even at big upscaling factors, recent improvements in the
also applied in a few other research works [27, 28]. One
Super Resolution Generative Adversarial Network
key challenge with such deep networks is the vanishing
(SRGAN) have been made [16]. The discriminator of
gradient issue. To get around the vanishing gradient issue,
ESRGAN is built in VGG-style using spectral
a dense skip connection was employed [29]. The sub-pixel
normalization and U-Net design [21, 22] in a recent
convolutional layer was first proposed for up-sampling in
SRGAN technique extension known as the Real Enhanced
CNNs [30]. For the same purpose, an improved sub-pixel
Super-Resolution Generative Adversarial Network (Real-
convolutional network was suggested [14]. The sub-pixel
ESRGAN) [23]. The training dynamics are stabilized, and
up-sampling layer fared better than the transpose
the capability of the discriminator is increased. Real-
convolution-based and interpolation-based up-sampling
ESRGAN attains improved visual performance, as a result
techniques. A channel attention mechanism was put forth
increasing its usefulness in practical situations. Hence, it
by Zhang et al. [31] to enhance the effectiveness of image
provides the motivation to use Real-ESRGAN, which
SR techniques. The authors incorporated channel-
reconstructs 2D real-world natural images with excellent
dependent features using local receptive fields. The
perceptual quality. As far as we are aware, no research has
method's drawback is that certain non-local traits may go
been done on the application of Real-ESRGAN to validate
unnoticed. To solve this issue, Dai et al. [32] employed both
single-image SR on medical images. This work's
non-local and local receptive fields for the purpose of
significant contributions are as follows:
feature extraction. The robustness of the features that were
• In this work, we specifically aim to use Real-ESRGAN
extracted determines how well the super-resolution
to enhance the resolution of medical images.
approach performs. Few authors adopted the ResNet34
• At large upscaling factors, edge and minute texture
architecture [33] due to its improved performance [34, 35].
information is retained.
Ahn et al. [36] employed a series of residual network blocks
• Our suggested method, which employs Residual-in-
using the same concept. A prominent drawback of
Residual Dense Blocks (RRDB) blocks in the generator and
traditional CNNs is their limited ability to recover texture
U-net in the discriminator, exhibits superior super-
characteristics from images when subjected to large
resolution performance on many medical imaging
upscaling factors. Chauhan et al. [37] presented a complete
modalities when compared to the state-of-the-art SR
methods currently in use.
2

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

review of deep learning strategies used for SR of single brain MRI imaging. In comparison to prior GAN-based
images. models, the scientists claimed a six-fold gain in speed and
The GAN approach is useful for increasing image spatial improved accuracy with the proposed network. Mohsen et
resolution. Generative Adversarial Networks (GAN) based al. [46] also achieved success in SR of brain MRI images
techniques have displayed considerable enhancement for two types of tumors, namely pituitary and glioma. Do
among deep learning-based SR methods [38–40]. The et al. [47] employed cycleGAN and ESRGAN for the
content-based or perceptual loss function, i.e., VGG19 loss, generation of HR MRI images possessing detailed textures.
is used by the GAN architecture to get over the drawbacks Their investigational findings were obtained at various
of CNN-based SR techniques. SRGAN, the pioneer GAN- resolution scales using both 3T and 7T MRI. In order to
based SR technique, was put forth in [16]. A ResNet34 produce HR MRI images, the authors in [48] utilized the
architecture serves as the SRGAN generator, extracting SRGAN-based approach [16] adapted to 3D convolutions.
features from images that are in LR. A sub-pixel They have investigated various techniques for the
convolution layer was utilized for upsampling. Wang et al. upsampling stage in order to reduce artifacts resulting from
[23] made further modifications to this procedure. They sub-pixel convolution layers. Few other recent research
took advantage of residual-in-residual dense blocks, that is, works have also explored image enhancement for MRI
multi-residual networks. The batch normalization layer was scans [49-51]. Lin et al. presented image enhancement for
also eliminated by the authors after each convolutional 3D MRI scans using deep GAN via Slice-Profile
layer. The traditional SRGAN design performed better Transformation SR [52]. Temiz et al. [53] utilized a CNN
once the batch normalization layer was removed. model for achieving SR of ultrasound images. The cycle-
Computational complexity increases when using several consistent Generative Model was used for SR in X-ray
residual blocks or nested residual blocks. Edge-enhanced Microscopy [54], which was inspired by the achievement
GAN was developed by Jiang et al. [41] for satellite image of CycleGAN. In order to ensure "cycle consistency" across
SR. The architecture consists of two subnetworks: in the the two domains (LR and HR) for unpaired image-to-image
first subnetwork, features are extracted, and HR features translation, their model employs two pairs of generator and
are obtained possessing sharp edges, known as ultra-dense discriminator blocks. SRGAN models have been
subnetwork (UDSN); in the second subnetwork, the extensively used for enhancing the chest X-rays [55-57].
extracted sharp edges are enhanced further, and artifacts are SR enhancement has been applied to lesion images to detect
eliminated which originated in the course of UDSN feature skin cancer and other diseases in recent works [58-61].
extraction process, second sub-network is known as edge GAN-based SR algorithms have been put into application
enhanced sub-network (EESN). for the upscaling of ultrasound imaging [62, 63]. The
Gu et al. [42] proposed SRGAN for medical images. current research, Real-ESRGAN for medical images
GAN-based approaches addressing SR for medical images presented here, has achieved further improvements.
include mDCSRN [43], Lesion-focussed GAN [44], and We summarize the advantages, limitations, and
ESRGAN [45]. Chen et al. [43] have applied a multi-level, techniques used by main models for image super-resolution
densely connected SR network (mDCSRN) to improve from existing literature in Table 1.

TABLE I
THE TECHNIQUES USED IN THE EXISTING LITERATURE, THEIR ADVANTAGES AND LIMITATIONS

Model/Method Techniques Used Advantages Limitations

SRCNN [25] - Three-layer CNN - Pioneer deep learning method for - Limited depth and complexity
- Bilinear interpolation for image SR - Restricted capacity to capture
up sampling - Outperforms conventional SR intricate features
methods
Deep Recursive CNN [26] - Recursive learning - Effective for extracting complex - Vanishing gradient issue
- Dense skip connection features
- Deeper network improves
performance
Sub-pixel Convolutional - Sub-pixel convolution for - Superior to transpose convolution - High computational cost
Network [30] up sampling and interpolation for up sampling
Residual Channel Attention - Channel attention - Enhances image SR by focusing - May overlook certain non-
Mechanism [31] mechanism with local on channel-dependent features local features
receptive fields for better
feature extraction
Second-order Attention - Non-local and local - Combines non-local and local - Increased complexity and
Network [32] receptive field combination receptive fields for better feature potential computational burden
extraction
ResNet-based SR [33] - ResNet34 architecture - Improved feature extraction and - High computational cost due
- Residual blocks image reconstruction to multiple residual blocks

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

GAN-based SR [16] - Generative Adversarial - Significant enhancement in image - High computational cost
Networks details and realism - Possible artifacts and
- VGG19 perceptual loss - Handles perceptual loss well instability in training
function
Edge-enhanced GAN [41] - Two subnetworks: UDSN - Improved sharpness and edge - Complex architecture
and EESN details in HR images increases computation
- Effective for satellite images - Risk of overfitting
MedSRGAN [42] - GAN-based SR for - Adapted specifically for medical - Still computationally
medical images images intensive
- Customized network - Enhanced clarity in medical SR - Limited generalizability
architecture tasks beyond medical domain
mDCSRN [43] - Multi-level densely - Improved speed and accuracy for - May not be optimal for other
connected network MRI SR medical imaging modalities
(mDCSRN) - Better performance in brain
- 3D convolution imaging
CycleGAN [47] - Cycle consistency loss - Effective for unpaired image-to- - High complexity
- Two pairs of generator image translation - Sensitive to training
and discriminator blocks - Preserves texture details instability
WSRGAN [62] - Wavelet based GAN - Enhanced resolution in X-ray and - Specific to certain medical
tailored for X-ray and ultrasound images imaging types
ultrasound - Improves diagnostic accuracy - Computationally expensive
- Enhanced CNN models without sampling loss

training introduces. The architecture of Real-ESRGAN is

III. PROPOSED SYSTEM presented in Fig. 2. In all convolutional layers, kernel size
This section explains the GAN model variant used by us to is 3, represented by the K variable; the number of feature
upscale the resolution of medical images, the dataset maps is 64, represented by the n variable, and a stride of 1,
description, training details with adjustments we made to represented by the s variable. With these modifications,
the model so that we could use it with various datasets of Real-ESRGAN is trained.
medical images, and the evaluation metrics.

A. ARCHITECTURE Of REAL-ESRGAN
The Real-ESRGAN is a generative-adversarial network
designed with 16 identical RRDB in the generator network
and a U-Net model in the discriminator network. The
generator used is the same as in the ESRGAN model, which is
a deep network with multiple RRDBs. Initial ESRGAN
architecture is expanded to achieve an SR of ×4 scale factor.
The pixel-unshuffle process is used to minimize the spatial
size and increase the channel size. As a result of this,
maximum computations are done in small resolution space,
hence employing less GPU memory and processing power.
The original discriminator architecture used in ESRGAN is no
longer appropriate because Real-ESRGAN attempts to handle FIGURE 2. Architecture of real-ESRGAN model.
a considerably bigger degradation space than ESRGAN [23].
More discriminative power is specifically needed by the B. DATASETS USED
discriminator in Real-ESRGAN for complex training outputs. Four publicly available datasets from Kaggle are used for all
Rather than differentiating between global styles, it must also experiments. To check the robustness of the proposed
generate precise gradient feedback for local textures. algorithm for the medical images, we have used four distinct
Additionally, the VGG-style discriminator in ESRGAN is medical image types as they represent the diversity of medical
improved by incorporating a U-Net architecture with skip conditions. These comprise the following: a brain MRI dataset
connections [23]. The features of the first block are added one (BraTS 2020), a dermoscopy dataset of skin cancer images, a
element at a time, while the features of the second block utilize 2D cardiac dataset of ultrasound images, and chest X-rays
skip connections. Skip connection aids in reducing the from the MIMIC-CXR dataset. Each dataset description is
vanishing gradient issue. Per-pixel feedback is given to the provided underneath.
generator by the U-Net as it produces the realness value for 1) BRAIN TUMOR MRI DATASET
each pixel. The training instability is also increased by the The BraTS 2020 dataset contains 285 MRI volumes with 152
complicated degradations and the U-Net structure. Thus, slice depths that are translated into 2D image slices. To fine-
spectral normalization regularization [64] is used to stabilize tune the pre-trained model, we use 257 MRI volumes, i.e.,
the training and reduce the overly sharp artifacts that GAN 257x152=39,064 of 2D MR image slices. Testing is done on
4

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

the final 28 MRI volumes, or 28x152=4,256 2D MR image loss, and GAN loss, and for training the discriminator, GAN
slices. The images are down-sampled by a factor of 4 in each loss was used.
x-y dimension to produce the LR images. The SR experiment We summed the overall loss data to train the generator by
for brain MRI images utilizes a batch size of 4. assigning a weight value of 1 to each of the various losses. In
2) SKIN CANCER DATASET
order to train the discriminator to differentiate between fake
The skin cancer dataset includes an enormous number of and real images, GAN loss was calculated for the real and fake
multi-source pigmented lesion images. The dataset comprises (produced from the generator) images. Perceptual loss is
a total of 33,126 dermoscopic images, from which 29,813 calculated using VGG19 weights. Equations 1, 2, and 3
images served as the training sets and the remaining images provide the mathematical equations for the aforementioned
served as testing sets. The images were down-sampled 4 times losses.
ℎ 𝑤
to serve as LR images. The skin cancer image SR experiment 1 2
𝐿𝑉𝐺𝐺 (𝐼 ′ , 𝐼) = ∑ ∑ (∅𝑖,𝑗 (𝐼 𝐻𝑅 )𝑥,𝑦 − ∅𝑖,𝑗 (𝐺(𝐼 𝐿𝑅 )𝑥,𝑦 ))
used a batch size of two. ℎ𝑤
𝑥=1 𝑦=1
3) CARDIAC ULTRASOUND IMAGES DATASET
Two- and four-chamber cardiac 2D ultrasound images are (1)
included in the CAMUS (Cardiac Acquisition for Multi- The generated image's VGG feature vector is denoted by
structure Ultrasound Segmentation) data collection. There are ∅𝑖,𝑗 (𝐺(𝐼 𝐿𝑅 )) image, and the VGG feature vector of the HR
five hundred cardiac ultrasound images in the publicly image is represented by ∅𝑖,𝑗 (𝐼 𝐻𝑅 ). The height and width of the
accessible dataset. We utilize a batch size of two. The number image are portrayed by the variables h and w, respectively.
of images employed for training is 400, and for testing is 100.
LR image is obtained by downsampling the images four times. 𝐿𝑔𝑒𝑛 = ∑𝑁 ′
𝑛=1 − log 𝐷(𝐼 ) (2)
4) CHEST X-RAY IMAGES DATASET
The Chest X-ray images from the MIMIC-CXR dataset are a Where the output of the discriminator is denoted by 𝐷(𝐼 ′ ).
huge collection of 371,920 images. For the training purpose,
227,943 images were used, and for evaluation, the rest of the 1 ′ 2
𝐿2 (𝐼 ′ , 𝐼) = ∑𝑖.𝑗.𝑘(𝐼𝑖,𝑗,𝑘 − 𝐼𝑖,𝑗,𝑘 ) (3)
ℎ𝑤𝑐
images were used. A factor of 4 is used to produce the LR
images. For the chest X-ray image SR experiment, a batch size ′
Where 𝐼𝑖,𝑗,𝑘 is the generated image and 𝐼𝑖,𝑗,𝑘 is the original
of four is employed.
image. The height, width, and the number of channels is
denoted by the variables h, w, and c, respectively. The absolute
C. TRAINING DETAILS
difference error (L1 loss), as displayed by (4), is also included
In this work, we employed transfer learning for training to
in our model along with other losses, which enhances the peak
enhance the performance and efficiency of our model.
signal-to-noise ratio. The drawback of L1 loss is that it
Transfer learning allows us to leverage pre-trained models that
oversmoothes the output. L2 loss is used in conjunction with
have been trained on large datasets, reducing the need for L1 loss to address this.
extensive training from scratch and enabling our model to
converge faster and perform better with limited data. The 1 ′
𝐿1 (𝐼 ′ , 𝐼) = ∑𝑖.𝑗.𝑘|𝐼𝑖,𝑗,𝑘 − 𝐼𝑖,𝑗,𝑘 | (4)
training procedure is outlined as follows. We selected a pre- ℎ𝑤𝑐
trained Real-ESRGAN model that has demonstrated strong
performance in image SR tasks. Real-ESRGAN is trained on Next, the total loss is computed using (5):
large-scale datasets such as DIV2K, which includes HR
images across diverse domains, making it a suitable choice for 𝑂𝑣𝑒𝑟𝑎𝑙𝑙 𝐿𝑜𝑠𝑠 = 𝐿1 (𝐼 ′ , 𝐼) + 𝐿2 (𝐼 ′ , 𝐼) + 𝐿𝑔𝑒𝑛 + 𝐿𝑉𝐺𝐺 (𝐼 ′ , 𝐼)
our application. The initial layers of the pre-trained model
were frozen to retain the learned feature representations of the (5)
previous model. The later layers, particularly those involved We have used these loss functions to evaluate the training
in high-level feature representation and decision-making, were performance. The training process loss graphs for the real-
fine-tuned to adapt to the specific characteristics of our ESRGAN model are displayed in Fig. 3. The blue curve
proposed model. Adam optimizer was used [65], known for its depicts the loss plot for the training phase of Real-ESRGAN
adaptive learning rate capabilities, to fine-tune the model model, and the red color depicts the loss plot for the validation
parameters. A lower learning rate of 1 × 10−4 is used initially phase of Real-ESRGAN model. The training and validation
to prevent significant deviation from the pre-trained weights. losses both start at a high value and decrease as the number of
A batch size of 16 was used for training. We employed a epochs increases. The validation loss being lower or close to
combination of pixel-wise loss (L1 loss), perceptual loss [66], the training loss in all three cases suggests that the model is
and GAN loss [16] to guide the training process. For training well-regularized and generalizes well to unseen data.
the generator, we combined pixel loss (L1 loss), perceptual

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

work was analyzed in comparison with a few other SR

approaches using the assessment metrics.
1) MEAN SQUARE ERROR
The most popular loss function metric utilized in deep learning
is mean square error (MSE), sometimes referred to as pixel
loss. It detects variations between the pixels of the ground-
truth image and the predicted image. Equation (6) shows the
MSE.
ℎ−1 𝑤−1
1 2
𝑀𝑆𝐸 = ∑ ∑(𝐼(𝑥, 𝑦) − 𝐼 ′ (𝑥, 𝑦))
ℎ𝑤
𝑥=0 𝑦=0

(6)
′
where 𝐼𝑥,𝑦 is the generated image, and 𝐼𝑥,𝑦 is the original
image. The height and width of the image is represented by the
variables h and w, respectively
2) PEAK SIGNAL‑TO‑NOISE RATIO
One way to determine the nobility of the rebuilt image is to
utilize the peak signal-to-noise ratio (PSNR). PSNR is widely
used as a standard benchmark in image processing tasks,
allowing for straightforward comparison of different models
or techniques. The PSNR measures the ratio amid a signal's
peak power and the noise that degrades its quality. PSNR is
used to quantify the difference between the original high-
resolution image (ground truth) and the super-resolved image
produced by the model. It helps in determining whether the SR
model is improving the raw image data compared to
conventional methods. PSNR provides a direct measure of
pixel-level accuracy, which is crucial in medical imaging
where small errors can lead to significant diagnostic
implications. High PSNR indicates that the super-resolved
image is very close to the original, meaning that fine details,
such as tissue boundaries and pathological features, are well
preserved. PSNR does not account for perceptual quality or
how a human observer would perceive the image. Two images
with the same PSNR might look different to the human eye,
especially in terms of texture or contrast, which is critical in
medical imaging. PSNR is represented mathematically in (7).

max (𝐼)2
𝑃𝑆𝑁𝑅 = 10. 𝑙𝑜𝑔10 ( ) (7)
𝑀𝑆𝐸

Where I denote the image. The PSNR will be higher when the
MSE is lower.
FIGURE 3. Loss function plots. (a) L1 loss, (b) perceptual loss, and (c)
GAN loss. 3) STRUCTURAL SIMILARITY INDEX MEASURE
SSIM (Structural Similarity Index Measure) is a perceptual
D. QUANTITATIVE EVALUATION METRICS metric that measures the similarity between two images based
It is crucial to evaluate the performance of the proposed model on their structural information. It compares the images in
using metrics that accurately reflect the quality of the terms of luminance, contrast, and structure, which are more
generated images. The widely used metrics in image aligned with human visual perception than mere pixel-wise
processing are Peak Signal-to-Noise Ratio (PSNR), Structural differences. SSIM focuses on maintaining the structural
Similarity Index (SSIM) and Mean Square Error (MSE). Each integrity of the image, which is paramount in medical imaging.
serves a distinct purpose in assessing the quality of super- For example, in medical diagnostics, the structure and shape
resolved images in the sensitive domain of medical imaging of anatomical features are crucial for identifying
where precision is critical. The performance of the suggested
6

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

abnormalities. SSIM ensures that these structures are the comparison with other cutting-edge techniques for the
preserved in the super-resolved images. Unlike PSNR, SSIM BraTS dataset. In our evaluation, the proposed model achieved
aligns better with human visual perception. This is especially a PSNR of 36.99 dB and a SSIM of 0.97, outperforming both
important in medical imaging, where the clinical utility of an the ESRGAN and SRGAN methods, which achieved a PSNR
image often depends on how clearly a radiologist can perceive of 36.4 dB, SSIM of 0.96 and PSNR of 33.55 dB, SSIM of
and interpret fine details and structures. SSIM provides a more 0.94, respectively. When compared to the SRGAN approach,
holistic measure of image quality by considering not just the suggested method increased the PSNR and SSIM by 3.44
pixel-wise accuracy but also how well the super-resolved dB and 3.2 percent, respectively, while the ESRGAN
image maintains the overall look and feel of the original technique outperformed the SRGAN method in terms of
image, which is important for accurate diagnosis. While SSIM PSNR and SSIM by 2.85 dB and 2.1 percent, respectively. The
is excellent for capturing structural similarity, it might higher value of PSNR suggests that our model produces brain
overlook small pixel-level differences that could be clinically MRI images that are closer to the original ones in terms of
relevant. Therefore, it is often used in conjunction with PSNR pixel-wise accuracy, while the increased SSIM indicates better
to provide a comprehensive evaluation of image quality. The preservation of structural details, which is crucial for
mathematical representation of SSIM is given by (8). maintaining the perceptual quality of the images. The MSE
value of real-ESRGAN is lower than the other two methods
(2𝜇1 𝜇2 +𝑐1)(2𝜎12 +𝑐2) and hence superior to them. These results demonstrate that our
𝑆𝑆𝐼𝑀(𝐼 ′ , 𝐼) = (8)
(𝜇12 +𝜇22 +𝑐1)(𝜎12 +𝜎22 +𝑐2)
approach effectively balances fidelity and perceptual quality,
making it well-suited for brain MRI image SR where both
Where 𝜎1 and 𝜎2 denote the variances of the original image I factors are critical.
and the reconstructed image 𝐼 ′ respectively, the equation is
stabilized by the two variables, c1, and c2, 𝜇1 , 𝜇2 is the average TABLE II
of the reconstructed image 𝐼 ′ , and the original image I, THE PSNR, SSIM, AND MSE OF SRGAN, ESRGAN AND REAL-ESRGAN
ALGORITHMS FOR BRAIN MRI IMAGES
respectively.
Using all the three metrics together provides a well-rounded Metric SRGAN ESRGAN Real ESRGAN
evaluation, ensuring that the proposed SR model not only PSNR 33.55 36.40 36.99
produces technically accurate images but also maintains the SSIM 0.94 0.96 0.97
structural and perceptual qualities necessary for effective MSE 16.99 14.79 13.19
medical diagnosis.
Fig. 4 provides the graphical comparison of SSIM, PSNR, and
IV. RESULTS ANALYSIS MSE of various models used for brain MRI images.
Within this section, we describe the outcomes of the real-
ESRGAN model we have employed to upscale the resolution
of medical images. The PSNR, SSIM, and MSE of the output B R A I N M R I I M A G ES
images were estimated based on the input images. Below is a
discussion of each dataset's results, along with a comparison PSNR SSIM MSE
with other state-of-the-art methods. 36.99
40 36.4
33.55
A. BRAIN MRI IMAGES 30
Fig. 5 displays SRGAN, ESRGAN and the Real-ESRGAN
Value

method's visual results on brain MRI images. The SR results 20 16.99 14.79 13.19
of one image are displayed in each row. It is visible from the
Fig. 5 that the SRGAN-based method produces slightly 10
0.94 0.96 0.97
improved results. The ESRGAN method improves the results 0
further. In the case of the real-ESRGAN technique, the results SRGAN ESRGAN Real-ESRGAN
are better than those of the other two techniques. These figures
Algorithm used
demonstrate that, for brain MRI scans, the real ESRGAN
approach produces clearer HR images than both SRGAN and
the ESRGAN methods. The PSNR, SSIM, and MSE values FIGURE 4. Graphical comparison of PSNR, SSIM, and MSE for brain
MRI images.
for the BraTS dataset are given in Table 2. The table illustrates

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

FIGURE 5. Results of SR for various images for BraTS dataset. Column (a) shows the ground truth images, column (b) has LR images, column (c) shows
the result of SRGAN, column (d) shows the output HR images for ESRGAN, and column (e) displays the output HR images for the real-ESRGAN method.
Every row displays the result of a single image.
approaches and far superior in comparison to the other two
B. DERMOSCOPIC IMAGES techniques. In terms of pixel-wise accuracy, the greater PSNR
The visual results of SRGAN, ESRGAN, and the Real- indicates that our model created SR dermoscopic images that
ESRGAN methods are shown in Fig. 6. The SR outcomes for are more like the original ones, and the higher SSIM shows
a single dermoscopic image are displayed in each row. As seen that structural features are better preserved, which is essential
in Fig. 6, the real-ESRGAN-based method produces improved for preserving the images' perceived quality. These findings
results. The HR images produced by real-ESRGAN are much show that our method successfully strikes a compromise
clearer than the HR images produced by the other two between fidelity and perceived quality, which makes it a good
methods. PSNR, SSIM, and MSE values for the skin cancer fit for dermoscopic image SR where both elements are crucial.
dataset, along with the comparison with SRGAN and
ESRGAN methods, are shown in Table 3. The SRGAN TABLE III
THE PSNR, SSIM, AND MSE OF SRGAN, ESRGAN, AND REAL-ESRGAN
technique obtained a PSNR of 35.31 dB and a SSIM of 0.93 ALGORITHMS FOR SKIN CANCER IMAGES
ESRGAN technique obtained a PSNR of 37.04 dB and a SSIM
of 0.95 whereas the proposed real-ESRGAN model achieved Metric SRGAN ESRGAN Real ESRGAN
a PSNR of 38.42 dB and a SSIM of 0.97 in our experiment. In PSNR 35.31 37.04 38.42
comparison to the SRGAN approach, the ESRGAN method SSIM 0.93 0.95 0.97
enhanced PSNR and SSIM by 1.73 dB and 2.2 percent, MSE 15.74 14.34 13.21
respectively whereas the real-ESRGAN method upscaled the
LR image by 3.11 dB and 4.3 percent, respectively. The MSE Fig. 7 provides the graphical comparison of SSIM, PSNR,
value of real-ESRGAN is again lowest than the other two and MSE of different GAN variants for dermoscopic images.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

FIGURE 6. Results of SR for various images for the ISIC skin cancer dataset. Column (a) shows the ground truth images, column (b) has LR images,
column (c) shows the result of SRGAN, column (d) shows the output HR images for ESRGAN, and column (e) displays the output HR images for the
real-ESRGAN method. Every row displays the result of a single image.

C. ULTRASOUND IMAGES
The output of the SRGAN, ESRGAN, and Real-ESRGAN
D E R M OS CO PI C I M A G ES algorithms can be viewed in Fig. 8. The SR outcomes for a
single ultrasound image are displayed in each row. As seen
PSNR SSIM MSE from the Fig. 8 improved results are produced by the real-
ESRGAN based method.
50
37.04 38.42 The PSNR, SSIM, and MSE values for the cardiac ultrasound
40 35.31
images from the CAMUS dataset are illustrated in Table 4.
30 The table also shows the comparison of these metrics with
Value

20 15.74 14.34 SRGAN and ESRGAN methods. The suggested real-

13.21
ESRGAN model achieved an average PSNR of 35.57 dB and
10 a SSIM of 0.86, surpassing the ESRGAN method which
0.93 0.95 0.97
0 achieved a PSNR of 33.96 dB and SSIM of 0.85, and the
SRGAN ESRGAN Real-ESRGAN SRGAN method, which recorded a PSNR of 31.55 dB and a
Algorithm used SSIM of 0.84. The enhancement by the ESRGAN method
concerning the SSIM and PSNR metric is 2.41 dB and 1.2
percent, respectively, while the enhancement by the real-
FIGURE 7. Graphical comparison of PSNR, SSIM, and MSE for
dermoscopic images.
ESRGAN method pertaining to PSNR and SSIM metric is
4.02 dB and 2.4 percent, respectively. Again, the MSE value

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

FIGURE 8. Results of SR for different images for CAMUS dataset. Column (a) shows the ground truth images, column (b) shows LR images, column
(c) shows the result of SRGAN, column (d) shows the output HR images for ESRGAN, and column (e) displays the output HR images for the real-
ESRGAN method. Every row displays the result of a single image.

is lowest for the real-ESRGAN method and, therefore, Fig. 10 provides the graphical comparison of SSIM, PSNR,
superior in comparison to the other two techniques. The higher and MSE of different GAN variants for ultrasound images.
PSNR indicates that our model produces SR ultrasound
images with greater pixel-wise accuracy relative to the original D. CHEST X-RAY IMAGES
image, while the improved SSIM reflects better preservation The visual performance of real-ESRGAN and other
of structural details, essential for maintaining perceptual comparative algorithms can be seen in Fig. 9. The SR results
quality. These results demonstrate that our approach of a single chest X-ray image are displayed in each row. Figure
successfully balances both the crucial aspects such as fidelity 9 illustrates how the real-ESRGAN-based approach yields
and perceptual quality, making it particularly well-suited for better outcomes. Compared to the HR images generated by the
ultrasound image super-resolution. other two techniques, the HR images generated by real-
ESRGAN are significantly clearer.
TABLE IV The PSNR, SSIM, and MSE values for the chest X-ray images
THE PSNR, SSIM, AND MSE OF SRGAN, ESRGAN, AND REAL-ESRGAN
from the MIMIC-CXR dataset are presented in Table 5. The
ALGORITHMS FOR ULTRASOUND IMAGES
comparison with different cutting-edge techniques on the
Metric SRGAN ESRGAN Real ESRGAN above-mentioned metrics for the MIMIC-CXR dataset is also
PSNR 31.55 33.96 35.57 shown in the table. During our evaluation, the real-ESRGAN
SSIM 0.84 0.85 0.86 model outperformed the ESRGAN and SRGAN technique
MSE 16.61 14.92 13.12 with a PSNR of 34.95 dB and an SSIM of 0.88, compared to

10

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

FIGURE 9. Results of SR on various images for MIMIC-CXR dataset. Column (a) shows the ground truth images, column (b) has LR images, column (c)
shows the result of SRGAN, column (d) shows the output HR images for ESRGAN, and column (e) displays the output HR images for the real-ESRGAN
method. Every row displays the result of one image.

33.81 dB and 0.86 for the ESRGAN method and 31.68 dB and
U LT R A S O UN D I M A G E S 0.85 for the SRGAN method. For the ESRGAN approach the
PSNR and SSIM metric values are increased by 2.13 dB and
SSIM PSNR MSE 1.2 percent, respectively in reference to the SRGAN approach.
For the real-ESRGAN approach, the PSNR and SSIM metric
40 values are increased by 3.27 dB and 3.5 percent, respectively,
33.96 35.57
35 31.55 in reference to the SRGAN approach. The MSE score of real-
30 ESRGAN is minimal and, hence, much better than the other
25 two techniques. Therefore, the higher the visual quality of an
Value

20 16.61 14.92 output image by a technique, the improved is its mean MSE,
13.12
15 PSNR, and SSIM values.
10 Better preservation of structural details is reflected in the
5 0.84 0.85 0.86 improved SSIM, while the higher PSNR shows that our model
0 generates super-resolved chest X-ray images with greater
SRGAN ESRGAN Real-ESRGAN pixel-wise accuracy. These results suggest that our technique
Algorithm used adequately balances fidelity with perceived quality, making it
particularly well-suited for chest X-ray imaging super-
resolution, where both elements are vital.
FIGURE 10. Graphical comparison of PSNR, SSIM, and MSE for
ultrasound images.

11

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

TABLE V F. ABLATION STUDY

THE PSNR, SSIM, AND MSE OF SRGAN, ESRGAN, AND REAL-ESRGAN In this section, we present an ablation study to evaluate the
ALGORITHMS FOR CHEST-X-RAYS IMAGES
contribution of various components of the Real-ESRGAN
Metric SRGAN ESRGAN Real ESRGAN model on the performance of SR in medical imaging.
Ablation study was performed in three different steps. In the
PSNR 31.68 33.81 34.95
SSIM 0.85 0.86 0.88 first step, the full Real-ESRGAN model is compared with a
MSE 17.51 14.14 12.96 variant where the RRDB is replaced by a simpler
architecture, such as a standard residual block. The PSNR
and SSIM metric values were decreased by 0.8 dB and 2
Fig. 11 provides the graphical comparison of SSIM, PSNR,
percent, respectively. The results demonstrate that RRDB
and MSE of different GAN variants for chest X-ray images.
significantly enhances the model's ability to recover fine
details in medical images, particularly in high-frequency
regions like edges and textures. The variant without RRDB
C H E ST X - R AY I M A G ES shows a noticeable decline in image sharpness and fails to
accurately reconstruct subtle anatomical features,
PSNR SSIM MSE
underscoring the importance of RRDB in maintaining image
40 fidelity. In the second step, we assess the impact of three
33.81 34.95
35 31.68 critical loss functions, L1 loss, perceptual loss, and GAN loss
30 on the performance of the model. We remove the L1 loss
from the training process and observe the effects on the
25
model’s performance. Without L1 loss, the model exhibits a
Value

20 17.15
14.14 12.96 noticeable decline in pixel-level accuracy. The resulting
15
images tend to deviate more from the ground truth in terms
10 of overall brightness and contrast, leading to reduced PSNR,
5 0.85 0.86 0.88 SSIM and increased MSE values. The PSNR and SSIM
0 metric values are decreased by 2 dB and 3 percent,
SRGAN ESRGAN Real-ESRGAN respectively, and MSE increased by 0.05. The absence of L1
Algorithm used loss also affects the consistency of reconstructed structures,
making the images less reliable for precise diagnostic
purposes. Next, we removed the perceptual loss from the
FIGURE 11. Graphical comparison of PSNR, SSIM, and MSE for chest X-
ray images. training process and found that omitting perceptual loss
results in images that, while retaining structural accuracy,
E. QUALITATIVE ASSESSMENT appear overly smooth and lack the texture detail that is often
Beyond quantitative metrics, the quality of the super- crucial in medical imaging. Fine anatomical features, such as
resolved images was also evaluated qualitatively by a small the texture of soft tissues or the boundaries of small lesions,
panel of experienced radiologists. The evaluation focused on are less distinct, making the images less useful for clinical
the clinical relevance of the images, including the clarity of interpretation despite their higher PSNR and SSIM values
anatomical details, the absence of artifacts, and overall compared to when perceptual loss is included. The PSNR
diagnostic utility. The feedback of radiologists is as follows: and SSIM values increased by 0.5 dB and 1 percent,
• Radiologists reported that the Real-ESRGAN generated respectively. Subsequently, we explored the effect of the
images were highly comparable to the original HR images, adversarial (GAN) loss by comparing the full Real-
particularly in maintaining critical diagnostic features such as ESRGAN model against a variant trained without the GAN
tissue boundaries, and lesions. component. We observed that absence of GAN loss results
in a significant drop in the realism and sharpness of the
• The model's ability to enhance image resolution without
generated images while maintaining the structural accuracy.
introducing clinically significant artifacts was noted as a key
In the third step, we examine the contribution of an improved
advantage over traditional upscaling methods.
discriminator that incorporates a U-Net architecture with
• However, some minor artifacts were observed in skip connections. For this purpose, the discriminator in real-
extremely complex regions, such as areas with dense tissue ESRGAN that incorporates a U-Net architecture with skip
overlap, suggesting room for further refinement. connections is replaced by a simple VGG style discriminator.
Therefore, radiologists observed that the images produced by The PSNR and SSIM metric values were decreased by 0.7
the Real-ESRGAN displayed a higher level of detail, dB and 2 percent, respectively. The U-Net-based
particularly in regions critical for diagnosis. This discriminator produces sharper images with better-defined
improvement was most evident in complex imaging textures. The incorporation of skip connections allows the
scenarios, such as the detection of small lesions or the model to better capture and reconstruct subtle details, leading
delineation of tissue boundaries. to images that are not only visually superior but also more
diagnostically useful.

12

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

Table 6 displays the results for all the ablation study's A. ADVANTAGES OF REAL-ESRGAN COMPARED TO
experiments. Fig. 12 presents the results of each experiment ESRGAN
of the ablation study visually. A detailed comparative analysis between our proposed
method (Real-ESRGAN) and ESRGAN technique is
TABLE VI presented in this section. The differences between Real-
ABLATION STUDY FOR THE PROPOSED METHOD USING DERMOSCOPIC ESRGAN and ESRGAN may seem subtle, they are
IMAGES
significant when applied in the context of the SR of medical
PSNR images where high precision and minimal artifacts are
Ablation experiment SSIM MSE
(dB)
critical. However, there are several key areas where Real-
Remove RRDB 37.62 0.95 14.01 ESRGAN offers advantages over ESRGAN.
Remove L1 Loss 36.42 0.94 13.26
Remove Perceptual Loss 37.92 0.96 13.37 1) ENHANCED IMAGE QUALITY
Remove GAN Loss 39.42 0.96 13.35
Remove U-net with skip Real-ESRGAN incorporates a more advanced discriminator
37.72 0.95 13.51
connections
framework, leading to superior detail preservation in the
Proposed Method 38.42 0.97 13.21
upscaled images. The comparative visual examples in the
Figs. 5, 6, 8, and 9 illustrate the finer details and textures that
One dermoscopic image has been displayed to illustrates the are preserved by Real-ESRGAN compared to ESRGAN.
results of ablation study experiment.
2) PRESERVATION OF STRUCTURAL FIDELITY

ESRGAN focuses on perceptual quality, sometimes at the

expense of structural fidelity. In medical imaging, however,
preserving the exact structure of anatomical features is
critical, as even small deviations can lead to incorrect
diagnoses. Real-ESRGAN places a greater emphasis on
maintaining structural fidelity, by incorporating network
designs that prioritize the accurate representation of edges,
textures, and other structural details. By preserving structural
fidelity more effectively, Real-ESRGAN ensures that super-
resolved medical images retain the anatomical accuracy
necessary for reliable diagnosis.
3) REDUCTION OF ARTIFACTS WITH ENHANCED GAN
FRAMEWORK

ESRGAN can occasionally produce perceptually noticeable

FIGURE 12. Results for ablation studies. (a) Represents the artifacts, which may not be appropriate for sensitive
corresponding HR ground truth image. (b) Represents the resultant SR
image without RRDB. (c) Represents the resultant SR image for applications like medical imaging where accuracy is crucial.
experiment without the L1 loss. (d) Represents the resultant SR image for The employment of Real-ESRGAN results in fewer artifacts
experiment after removing the perceptual loss function. (e) Represents
the resultant SR image for experiment without the adversarial loss. (f)
and better preservation of fine details, which are essential for
Represents the resultant SR image without the U-net architecture with the accurate interpretation of medical images. We have
skip connections in discriminator. included quantitative metrics, such as PSNR and SSIM,
demonstrating reduced artifact presence in real-ESRGAN
These findings underline the importance of the holistic results.
design of the Real-ESRGAN model for effective SR in
medical imaging. Therefore, the ablation study reveals that B. IMPACT OF ENHANCED IMAGES ON DIAGNOSTIC
each component of the Real-ESRGAN architecture ACCURACY AND CLINICAL DECISION-MAKING
contributes uniquely to its overall performance in medical
image SR. Removing any of these loss functions leads to The enhancement of medical images achieved through
trade-offs between different aspects of image quality, super-resolution techniques can have a significant impact on
highlighting the importance of their combined use to achieve diagnostic accuracy and clinical decision-making.
the best possible results in medical image SR. 1) IMPROVED VISUALIZATION OF ANATOMICAL
STRUCTURES
V. DISCUSSION
In this section, we discuss the advantages of Real- Enhanced image resolution can lead to better visualization of
ESRGAN compared to ESRGAN, the impact of enhanced fine details in anatomical structures. For instance, in brain
images on diagnostic accuracy and clinical decision- MRI scans, small lesions or micro-structural abnormalities
making, and the limitations of the proposed approach. that were previously undetectable in LR images may become
visible. This improved clarity can help in identifying
potential issues earlier, leading to more accurate diagnoses.

13

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

2) INCREASED DIAGNOSTIC CONFIDENCE and dermoscopic images), it might not generalize well to
other medical imaging modalities (CT scans). Due to the
Uncertainty in diagnoses can be reduced with the help of
unique characteristics of each modality, the model may need
higher-quality images. Confident decisions can be made with
to be adjusted or customized to optimize performance with
clearer images, hence reducing the likelihood of
distinct kinds of images.
misdiagnosis. For instance, the improved resolution in skin
imaging can help distinguish between benign and malignant 2) INTRODUCTION TO THE ARTIFACT
lesions in a better way, leading to more precise treatment SR methods may introduce artifacts that could change
plans. important diagnostic characteristics. This is especially
concerning for medical imaging because it's critical that little
3) REDUCTION IN THE NEED FOR ADDITIONAL
IMAGING
details be accurately depicted. It is imperative to assess the
clinical significance of these artifacts.
When the initial images are of higher quality, there may be 3) COMPUTATIONAL COMPLEXITY
less need for follow-up imaging studies to confirm findings. Like other deep learning models, Real-ESRGAN may incur
This not only reduces the burden on patients (in terms of significant computational costs. This could restrict its use in
time, cost, and exposure to additional radiation) but also situations with constrained computational capabilities, such
streamlines the diagnostic process. For instance, in X-ray or as remote clinics or mobile health units.
ultrasound imaging, the ability to capture more detail in a
single image could negate the need for further exploratory 4) TRAINING DATA LIMITATIONS
imaging. The model's performance is heavily influenced by the quality
and volume of training data. A small or biased training
4) ENHANCED DETECTION OF SUBTLE PATHOLOGIES dataset can cause the model to perform poorly on unseen or
Subtle pathological changes can be revealed with super diverse data, which could result in overfitting or
resolved images which might go unnoticed otherwise. This underperformance in real-world applications.
is especially significant for diseases that are still in their early 5) CLINICAL ADOPTION AND VALIDATION
stages, where early detection is essential for effective Although the model has demonstrated potential in research
treatment. For instance, enhanced images could enable settings, it requires comprehensive validation in clinical
earlier intervention and better patient outcomes in the case of contexts to ensure robustness and reliability. The acceptance
early-stage tumors or vascular anomalies. of AI-driven solutions in medical practice also depends on
5) SUPPORT FOR ADVANCED DIAGNOSTIC TOOLS regulatory approvals and integration into existing
workflows.
The performance of advanced diagnostic tools, such as
Computer-Aided Detection (CAD) systems can be improved VI. CONCLUSION AND FUTURE WORK
with enhanced images. For these systems to effectively The feasibility of applying an advanced deep learning
detect and classify anomalies, high-quality input data is approach, real-ESRGAN, to upscale the resolution of medical
required. By providing these systems with superior image images was investigated in this paper. The findings reveal that
quality, the likelihood of accurate automated analysis the real-ESRGAN model we have used improved the visual
increases, which can be used to support clinicians in their quality of the images from four publicly available datasets,
decision-making processes. namely the ISIC skin cancer dataset, the CAMUS dataset, the
6) IMPACT ON CLINICAL DECISION-MAKING BraTS dataset, and the MIMIC-CXR dataset. The HR images
reconstructed from LR images were assessed using three
Improving patient care is the ultimate objective of image
metrics viz. MSE, PSNR, and SSIM. The results depict that
quality improvement. Better-informed treatment decisions
the performance of the Real-ESRGAN model exceeded the
may result from clearer images. For instance, high-resolution
SRGAN and ESRGAN model in upscaling the images by a
brain pictures can be used to plan neurosurgery by giving
vital information regarding tumor boundaries, their factor of 4. The reconstructed images were perceptually and
proximity to critical brain structures, and other aspects that qualitatively better in comparison to the other two
influence surgical strategy. This can lead to more precise approaches.
surgeries, reduced complication rates, and better patient While the improvements in image quality are evident, the true
outcomes. value of these enhancements lies in their potential to improve
diagnostic accuracy and clinical decision-making. By
C. LIMITATIONS enabling earlier and more accurate detection of pathologies,
The potential limitations of the study are listed as below: reducing the need for additional imaging, and supporting more
confident clinical decisions, enhanced images can
1) GENERALIZATION TO DIVERSE MEDICAL
MODALITIES significantly contribute to better patient outcomes. Future
Although Real-ESRGAN works well for certain medical studies should aim to quantify these benefits through clinical
image types (such as MRI scans, Ultrasound, Chest X-ray trials and real-world validations, providing concrete evidence

14

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

of the impact of image enhancement technologies on collection of images in terms of disease and imaging modality.
healthcare. This approach could include collaboration with medical
Our future efforts will focus on doing a more thorough institutes to collect and interpret large-scale datasets. Methods
evaluation of the perceptual quality, involving other can be investigated to detect and mitigate artifacts introduced
quantitative and qualitative evaluations. Additionally, the during the super-resolution process. This could entail post-
model's capacity for training efficiency can be improved. processing procedures intended to improve the output. The
Lightweight versions of Real-ESRGAN can be developed that primary constraint on image SR is that the majority of current
maintain high performance while consuming less computing models only aim to target upscaling values by a factor of 4 or
power, particularly in environments with limited resources. below. Thus, further research is required to create a model that
Further, the generalizability of the model would be enhanced targets 8× or perhaps higher upscaling factors.
by enlarging the training dataset to contain a more varied
[15] Lim, B., Son, S., Kim, H., Nah, S., & Mu Lee, K. (2017). Enhanced
deep residual networks for single image super-resolution.
REFERENCES In Proceedings of the IEEE Conference on Computer Vision and
[1] Tanno, R., Worrall, D. E., Ghosh, A., Kaden, E., Sotiropoulos, S. N., Pattern Recognition Workshops (pp. 136-144).
Criminisi, A., & Alexander, D. C. (2017). Bayesian image quality [16] Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta,
transfer with CNNs: exploring uncertainty in dMRI super-resolution. A., ... & Shi, W. (2017). Photo-realistic single image super-resolution
In Medical Image Computing and Computer Assisted Intervention− using a generative adversarial network. In Proceedings of the IEEE
MICCAI 2017: 20th International Conference, Quebec City, QC, conference on computer vision and pattern recognition (pp. 4681-
Canada, September 11-13, 2017, Proceedings, Part I 20 (pp. 611-619). 4690).
Springer International Publishing. [17] Xiong, Y., Guo, S., Chen, J., Deng, X., Sun, L., Zheng, X., & Xu, W.
[2] Connolly, J. L., Schnitt, S. J., Wang, H. H., Longtine, J. A., Dvorak, A., (2020). Improved SRGAN for remote sensing image super-resolution
& Dvorak, H. F. (2003). Tumor structure and tumor stroma across locations and sensors. Remote Sensing, 12(8), 1263.
generation. Cancer medicine. [18] Zhao, Y., Chen, Z., Gao, X., Song, W., Xiong, Q., Hu, J., & Zhang, Z.
[3] Yang, H., Wang, Z., Liu, X., Li, C., Xin, J., & Wang, Z. (2023). Deep (2021). Plant disease detection using generated leaves based on
learning in medical image super-resolution: a review. Applied DoubleGAN. IEEE/ACM Transactions on Computational Biology and
Intelligence, 1-26. Bioinformatics, 19(3), 1817-1826.
[4] Li, Y., Sixou, B., & Peyrin, F. (2021). A review of the deep learning [19] Fleet, D., Pajdla, T., Schiele, B., & Tuytelaars, T. (2014). Learning a
methods for medical image super-resolution problems. Irbm, 42(2), deep convolutional network for image Super-Resolution Computer
120-133. Vision-ECCV 2014 European Conference on Computer Vision
[5] Rohith, G., & Kumar, L. S. (2021). Paradigm shifts in super-resolution (ECCV), pp 184–199.
techniques for remote sensing applications. The Visual [20] Yang, F., Yang, H., Fu, J., Lu, H., & Guo, B. (2020). Learning texture
Computer, 37(7), 1965-2008. transformer network for image super-resolution. In Proceedings of the
[6] Hayat, M., Armvith, S., & Achakulvisut, T. (2023, October). Combined IEEE/CVF conference on computer vision and pattern recognition (pp.
channel and spatial attention-based stereo endoscopic image super- 5791-5800).
resolution. In TENCON 2023-2023 IEEE Region 10 Conference [21] Schonfeld, E., Schiele, B., & Khoreva, A. (2020). A u-net based
(TENCON) (pp. 920-925). IEEE. discriminator for generative adversarial networks. In Proceedings of
[7] Hayat, M., & Aramvith, S. (2024). Saliency-Aware Deep Learning the IEEE/CVF conference on computer vision and pattern
Approach for Enhanced Endoscopic Image Super-Resolution. IEEE recognition (pp. 8207-8216).
Access. [22] Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net: Convolutional
[8] Hayat, M., & Aramvith, S. (2024). E-SEVSR-Edge Guided Stereo networks for biomedical image segmentation. In Medical Image
Endoscopic Video Super-Resolution. IEEE Access. Computing and Computer-Assisted Intervention–MICCAI 2015: 18th
[9] Nandal, P., & Pahal, S. (2024, February). Upscaling Images Using International Conference, Munich, Germany, October 5-9, 2015,
ESRGAN with Perceptual Loss. In International Conference On Proceedings, Part III 18 (pp. 234-241). Springer International
Innovative Computing And Communication (pp. 441-450). Singapore: Publishing.
Springer Nature Singapore. [23] Wang, X., Xie, L., Dong, C., & Shan, Y. (2021). Real-esrgan: Training
[10] El-Shafai, W., Ali, A. M., El-Nabi, S. A., El-Rabaie, E. S. M., & Abd real-world blind super-resolution with pure synthetic data.
El-Samie, F. E. (2024). Single image super-resolution approaches in In Proceedings of the IEEE/CVF international conference on computer
medical images based-deep learning: a survey. Multimedia Tools and vision (pp. 1905-1914).
Applications, 83(10), 30467-30503. [24] Abdou, M. A. (2022). Literature review: Efficient deep neural networks
[11] Yu, J., Lin, Z., Yang, J., Shen, X., Lu, X., & Huang, T. S. (2018). techniques for medical image analysis. Neural Computing and
Generative image inpainting with contextual attention. In Proceedings Applications, 34(8), 5791-5812.
of the IEEE conference on computer vision and pattern recognition (pp. [25] Dong, C., Loy, C. C., He, K., & Tang, X. (2015). Image super-
5505-5514). resolution using deep convolutional networks. IEEE transactions on
[12] Dong, C., Loy, C. C., He, K., & Tang, X. (2014). Learning a deep pattern analysis and machine intelligence, 38(2), 295-307.
convolutional network for image super-resolution. In Computer [26] Kim, J., Lee, J. K., & Lee, K. M. (2016). Deeply recursive
Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, convolutional network for image super-resolution. In Proceedings of
September 6-12, 2014, Proceedings, Part IV 13 (pp. 184-199). Springer the IEEE conference on computer vision and pattern recognition (pp.
International Publishing. 1637-1645).
[13] Dong, C., Loy, C. C., & Tang, X. (2016). Accelerating the super- [27] Tai, Y., Yang, J., & Liu, X. (2017). Image super-resolution via deep
resolution convolutional neural network. In Computer Vision–ECCV recursive residual network. In Proceedings of the IEEE conference on
2016: 14th European Conference, Amsterdam, The Netherlands, computer vision and pattern recognition (pp. 3147-3155).
October 11-14, 2016, Proceedings, Part II 14 (pp. 391-407). Springer [28] Li, Y., Zhang, L., Dingl, C., Wei, W., & Zhang, Y. (2018, September).
International Publishing. Single hyperspectral image super-resolution with grouped deep
[14] Shi, W., Caballero, J., Huszár, F., Totz, J., Aitken, A. P., Bishop, R., ... recursive residual network. In 2018 IEEE Fourth International
& Wang, Z. (2016). Real-time single image and video super-resolution Conference on Multimedia Big Data (BigMM) (pp. 1-4). IEEE.
using an efficient sub-pixel convolutional neural network. [29] Chen, C., & Qi, F. (2018, October). Single image super-resolution using
In Proceedings of the IEEE conference on computer vision and pattern deep CNN with dense skip connections and inception-resnet. In 2018
recognition (pp. 1874-1883).
15

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

9th International Conference on Information Technology in Medicine [49] Ebner, M., Patel, P., Atkinson, D., Caselton, L., Taylor, S., Bainbridge,
and Education (ITME) (pp. 999-1003). IEEE. A., ... & Vercauteren, T. (2019, May). Reconstruction-based super-
[30] Shi, W., Caballero, J., Theis, L., Huszar, F., Aitken, A., Ledig, C., & resolution for high-resolution abdominal MRI: a preliminary study.
Wang, Z. (2016). Is the deconvolution layer the same as a convolutional International Society for Magnetic Resonance in Medicine (ISMRM).
layer?. arXiv preprint arXiv:1609.07009. [50] Deka, B., Datta, S., Mullah, H. U., & Hazarika, S. (2020). Diffusion-
[31] Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., & Fu, Y. (2018). Image weighted and spectroscopic MRI super-resolution using sparse
super-resolution using very deep residual channel attention networks. representations. Biomedical Signal Processing and Control, 60,
In Proceedings of the European conference on computer vision 101941.
(ECCV) (pp. 286-301). [51] Li, Y., Song, B., Guo, J., Du, X., & Guizani, M. (2019). Super-
[32] Dai, T., Cai, J., Zhang, Y., Xia, S. T., & Zhang, L. (2019). Second-order resolution of brain MRI images using overcomplete dictionaries and
attention network for single image super-resolution. In Proceedings of non-local similarity. IEEE Access, 7, 25897-25907.
the IEEE/CVF conference on computer vision and pattern [52] Lin, J., Miao, Q., Surawech, C., Raman, S. S., Zhao, K., Wu, H. H., &
recognition (pp. 11065-11074). Sung, K. (2023). High-Resolution 3D MRI With Deep Generative
[33] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for Networks via Novel Slice-Profile Transformation Super-Resolution.
image recognition. In Proceedings of the IEEE conference on computer IEEE Access.
vision and pattern recognition (pp. 770-778). [53] Temiz, H., & Bilge, H. S. (2020). Super resolution of B-mode
[34] Han, W., Chang, S., Liu, D., Yu, M., Witbrock, M., & Huang, T. S. ultrasound images with deep learning. IEEE Access, 8, 78808-78820.
(2018). Image super-resolution via dual-state recurrent networks. [54] Raghunath, A., Wagner, F., Thies, M., Gu, M., Pechmann, S., Aust, O.,
In Proceedings of the IEEE conference on computer vision and pattern ... & Maier, A. (2023, June). Unsupervised Super Resolution in X-ray
recognition (pp. 1654-1663). Microscopy using a Cycle-consistent Generative Model. In BVM
[35] Lai, W. S., Huang, J. B., Ahuja, N., & Yang, M. H. (2017). Deep Workshop (pp. 76-81). Wiesbaden: Springer Fachmedien Wiesbaden.
laplacian pyramid networks for fast and accurate super-resolution. [55] Moran, M. B., Faria, M. D., Giraldi, G. A., Bastos, L. F., & Conci, A.
In Proceedings of the IEEE conference on computer vision and pattern (2021). Using super-resolution generative adversarial network models
recognition (pp. 624-632). and transfer learning to obtain high resolution digital periapical
[36] Ahn, N., Kang, B., & Sohn, K. A. (2018). Fast, accurate, and radiographs. Computers in biology and medicine, 129, 104139.
lightweight super-resolution with cascading residual network. [56] Xu, L., Zeng, X., Huang, Z., Li, W., & Zhang, H. (2020). Low-dose
In Proceedings of the European conference on computer vision chest X-ray image super-resolution using generative adversarial nets
(ECCV) (pp. 252-268). with spectral normalization. Biomedical Signal Processing and
[37] Chauhan, K., Patel, S. N., Kumhar, M., Bhatia, J., Tanwar, S., Control, 55, 101600.
Davidson, I. E., ... & Sharma, R. (2023). Deep learning-based single- [57] Bai, H., Zhou, X., Zhao, Y., Zhao, Y., & Han, Q. (2023). Soil CT image
image super-resolution: a comprehensive review. IEEE Access, 11, quality enhancement via an improved super-resolution reconstruction
21811-21830. method based on GAN. Computers and Electronics in Agriculture, 213,
[38] Shamsolmoali, P., Zareapoor, M., Granger, E., Zhou, H., Wang, R., 108177.
Celebi, M. E., & Yang, J. (2021). Image synthesis with adversarial [58] Abd El-Fattah, I., Ali, A. M., El-Shafai, W., Taha, T. E., & Abd El-
networks: A comprehensive survey and case studies. Information Samie, F. E. (2023). Deep-learning-based super-resolution and
Fusion, 72, 126-146. classification framework for skin disease detection
[39] Iqbal, T., & Ali, H. (2018). Generative adversarial network for medical applications. Optical and Quantum Electronics, 55(5), 427.
images (MI-GAN). Journal of medical systems, 42, 1-11. [59] Mukadam, S. B., & Patil, H. Y. (2023). Skin Cancer Classification
[40] Yuan, Y., Liu, S., Zhang, J., Zhang, Y., Dong, C., & Lin, L. (2018). Framework Using Enhanced Super Resolution Generative Adversarial
Unsupervised image super-resolution using cycle-in-cycle generative Network and Custom Convolutional Neural Network. Applied
adversarial networks. In Proceedings of the IEEE conference on Sciences, 13(2), 1210.
computer vision and pattern recognition workshops (pp. 701-710). [60] Vasamsetti, M., Kaja, P., Putta, S., & Kumar, R. (2023). Combining
[41] Jiang, K., Wang, Z., Yi, P., Wang, G., Lu, T., & Jiang, J. (2019). Edge- Super-Resolution GAN and DC GAN for Enhancing Medical Image
enhanced GAN for remote sensing image super-resolution. IEEE Generation: A Study on Improving CNN Model Performance. In GANs
Transactions on Geoscience and Remote Sensing, 57(8), 5799-5812. for Data Augmentation in Healthcare (pp. 187-205). Cham: Springer
[42] Gu, Y., Zeng, Z., Chen, H., Wei, J., Zhang, Y., Chen, B., ... & Lu, Y. International Publishing.
(2020). MedSRGAN: medical images super-resolution using [61] Hernandez, R. M., Ragasa, A. A., Macatangay, L., Palad, L., Mendez,
generative adversarial networks. Multimedia Tools and C., & Mendez, C. (2023, October). Enhanced Super-Resolution
Applications, 79, 21815-21840. Generative Adversarial Networks (ESRGAN) on Melanoma Skin
[43] Chen, Y., Shi, F., Christodoulou, A. G., Xie, Y., Zhou, Z., & Li, D. Cancer Detection. In Proceedings of the 8th International Conference
(2018, September). Efficient and accurate MRI super-resolution using on Sustainable Information Engineering and Technology (pp. 137-
a generative adversarial network and 3D multi-level densely connected 142).
network. In International conference on medical image computing and [62] Lyu, Y., Xu, Y., Yang, C., Jiang, X., Zhao, X., & Zhu, X. (2024).
computer-assisted intervention (pp. 91-99). Cham: Springer WSRGAN: A wavelet-based GAN for super-resolution of plane-wave
International Publishing. ultrasound images without sampling loss. Engineering Applications of
[44] Zhu, J., Yang, G., & Lio, P. (2019, April). How can we make GAN Artificial Intelligence, 127, 107384.
perform better in single medical image super-resolution? A lesion [63] Lyu, Y., Jiang, X., Xu, Y., Hou, J., Zhao, X., & Zhu, X. (2023). ARU-
focused multi-scale approach. In 2019 IEEE 16th International GAN: U-shaped GAN based on Attention and Residual connection for
Symposium on Biomedical Imaging (ISBI 2019) (pp. 1669-1673). super-resolution reconstruction. Computers in Biology and
IEEE. Medicine, 164, 107316.
[45] Bing, X., Zhang, W., Zheng, L., & Zhang, Y. (2019). Medical image [64] Miyato, T., Kataoka, T., Koyama, M., & Yoshida, Y. (2018). Spectral
super resolution using improved generative adversarial networks. IEEE normalization for generative adversarial networks. arXiv preprint
Access, 7, 145030-145038. arXiv:1802.05957.
[46] Mohsen, S., Ali, A. M., El-Rabaie, E. S. M., ElKaseer, A., Scholz, S. [65] Zhang, Z. (2018, June). Improved adam optimizer for deep neural
G., & Hassan, A. M. A. (2023). Brain Tumor Classification Using networks. In 2018 IEEE/ACM 26th international symposium on quality
Hybrid Single Image Super-Resolution Technique with of service (IWQoS) (pp. 1-2). IEEE.
ResNext101_32x8d and VGG19 Pre-Trained Models. IEEE Access. [66] Johnson, J., Alahi, A., & Fei-Fei, L. (2016). Perceptual losses for real-
[47] Do, H., Bourdon, P., Helbert, D., Naudin, M., & Guillevin, R. (2021). time style transfer and super-resolution. In Computer Vision–ECCV
7T MRI super-resolution with Generative Adversarial 2016: 14th European Conference, Amsterdam, The Netherlands,
Network. Electronic Imaging, 33, 1-7. October 11-14, 2016, Proceedings, Part II 14 (pp. 694-711). Springer
[48] Sánchez, I., & Vilaplana, V. (2018). Brain MRI super-resolution using International Publishing.
3D generative adversarial networks. arXiv preprint arXiv:1812.11440.

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3497002

Priyanka Nandal received the Ph. D series including 110 papers accepted and published in TOP
degree from Indian Institute of SCI indexed Journals. He also has 7 published Patents and 1
Technology Delhi, India in the year 2017. granted German Patent to his credit. Additionally, He has co-
She is currently working as an Associate authored and editing around 50 books. He has also served as a
Professor in Maharaja Surajmal Institute keynote speaker, resource person and given several invited
of Technology affiliated to Guru Gobind talks.
Singh Indraprastha University, Delhi,
India. She has 20 years of academic Placido Rogério Pinheiro is Ph.D.
experience and 11 years of parallel research experience. She degree in systems engineering and
has numerous research publications including many SCIE- computing degree by the Federal
indexed journal articles to her credit. Her research interests University of Rio de Janeiro, Rio de
include deep learning, computer vision, and high-speed Janeiro, Brazil. Currently, he is a Titular
computing. Professor of University of Fortaleza,
Retired Associate Professor of the State
Sudesh Pahal is working as an Associate University of Ceara and Research
Professor in Electronics and Productivity Scholarship (PQ) from the
Communication Engineering Department National Council for Scientific and Technological
at Maharaja Surajmal Institute of Development (CNPq) since 2008. He has experience in
Technology, Delhi (affiliated to GGSIP industrial processes modeling applying mathematical
university), India. She has completed her programming and multicriteria. His academic formation
B.E., M.E. and Ph.D in Electronics and allows publishing in applied mathematics area, with an
Communication Engineering. She has a emphasis on discrete and combinatorics mathematics,
teaching and research experience of more than 17 years. She working mainly in mathematical programming and
has Received Academic Excellent Performance Certificate for multicriteria. He has published in international periodicals
outstanding performance in academics and administration and with highlights for his citations.
Research Excellence Certificate in recognition of an
outstanding contribution to the quality of research. She is also
awarded with certificate of excellent mentorship for guiding
SIH winners and National Innovation Contest. She is also
Certified as Innovation Ambassador and mentor of start-ups
selected at national level by MHRD. She has published o2
books, 3 patents and more than 50 quality research papers in .
SCI/Scopus/ reputed journals and conferences. She has
delivered expert lectures in the field of Mobile
Communications, Wireless Networks, mobility management,
machine learning and vehicular networks, Intelligent
Transportation systems, Entrepreneurship guide.

Ashish Khanna [IEEE Member’2019,

Senior Member’2020, MACM] has
expertise in Teaching,
Entrepreneurship, and Research &
Development with specialization in
Computer Science Engineering
Subjects. He is IEEE EXECOM
MEMBER 2022,23,24 and treasurer in
IEEE DELHI COMSOC CHAPTER
AND MEMBER IEEE DELHI SECTION. He is part of
worlds leading 2% research community list from
STANFORD UNIVERSITY, USA 2022, 2023. He is also part
of top researcher community from AD research community
and research.com 2022 group. He has done post doc from
INATEL BRAZIL AND University of Valladolid, Spain,
Europe. He received his Ph.D. degree from National Institute
of Technology, Kurukshetra in March 2017. He has around
200 accepted and published research papers and book chapters
in reputed SCI, Scopus journals, conferences and reputed book

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4