Received 3 February 2024, accepted 26 February 2024, date of publication 1 March 2024, date of current version 8 March 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3372588

Pneumonia Detection Using Chest Radiographs

With Novel EfficientNetV2L Model
MUDASIR ALI1 , MOBEEN SHAHROZ2 , UROOJ AKRAM2 , MUHAMMAD FAHEEM MUSHTAQ 2,

STEFANÍA CARVAJAL ALTAMIRANDA3,4,5 , SILVIA APARICIO OBREGON3,6,7 ,

ISABEL DE LA TORRE DÍEZ 8 , AND IMRAN ASHRAF 9
1 Department of Computer Science, The Islamia University of Bahawalpur, Bahawalpur, Punjab 63100, Pakistan
2 Department of Artificial Intelligence, The Islamia University of Bahawalpur, Bahawalpur, Punjab 63100, Pakistan
3 Universidad Europea del Atlántico, 39011 Santander, Spain
4 Universidade Internacional do Cuanza, Cuito, Bié, Angola
5 Fundación Universitaria Internacional de Colombia, Bogotá, Colombia
6 Universidad Internacional Iberoamericana, Campeche 24560, Mexico
7 Universidad Internacional Iberoamericana Arecibo, Arecibo, PR 00613, USA
8 Department of Signal Theory, Communications and Telematics Engineering, University of Valladolid, 47011 Valladolid, Spain
9 Department of Information and Communication Engineering, Yeungnam University, Gyeongsan-si 38541, South Korea

Corresponding authors: Imran Ashraf ([email protected]) and Muhammad Faheem Mushtaq ([email protected])
This work was supported by European University of Atlantic.

ABSTRACT Pneumonia is a potentially life-threatening infectious disease that is typically diagnosed

through physical examinations and diagnostic imaging techniques such as chest X-rays, ultrasounds
or lung biopsies. Accurate diagnosis is crucial as wrong diagnosis, inadequate treatment or lack of
treatment can cause serious consequences for patients and may become fatal. The advancements in deep
learning have significantly contributed to aiding medical experts in diagnosing pneumonia by assisting
in their decision-making process. By leveraging deep learning models, healthcare professionals can
enhance diagnostic accuracy and make informed treatment decisions for patients suspected of having
pneumonia. In this study, six deep learning models including CNN, InceptionResNetV2, Xception, VGG16,
ResNet50 and EfficientNetV2L are implemented and evaluated. The study also incorporates the Adam
optimizer, which effectively adjusts the epoch for all the models. The models are trained on a dataset
of 5856 chest X-ray images and show 87.78%, 88.94%, 90.7%, 91.66%, 87.98% and 94.02% accuracy
for CNN, InceptionResNetV2, Xception, VGG16, ResNet50 and EfficientNetV2L, respectively. Notably,
EfficientNetV2L demonstrates the highest accuracy and proves its robustness for pneumonia detection.
These findings highlight the potential of deep learning models in accurately detecting and predicting
pneumonia based on chest X-ray images, providing valuable support in clinical decision-making and
improving patient treatment.

INDEX TERMS Pneumonia detection, transfer learning, efficientnetv2l, data augmentation, chest X-rays.

I. INTRODUCTION might result in mortality from pneumonia. Pneumonia is an

Pneumonia is an infectious disease that causes inflammation infectious disease caused by fungus, bacteria and viruses that
of the tissues in one or both lungs. Over one million people inflame the air sacs in one or both lungs. Lungs infection
are hospitalized in the United States each year due to harms the pulmonary alveoli which are little balloon-shaped
pneumonia [1]. Fortunately, pneumonia can be treated with bags at the bottom of the bronchioles. Pneumonia is classified
medications such as antibiotics and antivirals. Early diagnosis into numerous forms, including mycoplasma pneumonia,
and treatment are essential to prevent complications that viral pneumonia, bacterial pneumonia and others. Bacterial
pneumonia is caused by bacteria or fungus. Bacterial
The associate editor coordinating the review of this manuscript and pneumonia causes a variety of symptoms including physical
approving it for publication was Khursheed Aurangzeb. weakness, old age, disease, inadequate nutrition and a
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 34691
 M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

weakened immune system [2]. It is risky for people of subsequently employing a range of classifiers to refine the
all ages but especially dangerous for smokers, alcoholics, results [7].
recent surgical patients, asthmatics, virus infections and those Detecting pneumonia accurately from chest X-ray images
with weakened immune systems. Viruses that cause viral is a complex problem in medical diagnostics. The challenge
pneumonia such as the flu, account for nearly one-third lies in identifying the intricate and subtle patterns associ-
of all pneumonia cases. When a person is attacked by ated with pneumonia. Traditional methods rely on manual
viral pneumonia, the risk of developing bacterial pneumonia interpretation by radiologists which can be subjective and
increases, and more likely to develop bacterial pneumonia as time-consuming [8]. Deep learning-based approaches offer
well [3]. promise in automating the detection process. One major
Pneumonia is a common and potentially life-threatening challenge is the scarcity of annotated data for training deep
respiratory infection that poses a significant global health learning models. Annotated datasets of chest X-ray images
burden [4]. The timely and accurate detection of pneumonia with pneumonia cases are often limited and imbalanced,
is critical for effective treatment and patient care. Deep making it difficult for models to generalize effectively [9].
learning models use multiple layers of artificial neural Additionally, variations in imaging conditions including
networks to provide cutting-edge solutions in a variety of equipment positioning and image quality further complicate
fields including speech recognition, language translation, the task. Overcoming these challenges and developing a
image processing and others. Over the past few years, the robust deep learning model to precisely identify pneumonia
utilization of deep learning methods has surfaced as highly from chest X-rays is crucial for enhancing diagnostic
encouraging methods for the automation of pneumonia diag- efficiency and improving patient outcomes in clinical set-
nosis. Convolutional neural networks (CNNs) and recurrent tings. Deep learning methodologies effectively tackle the
neural networks (RNNs) have showcased their potential challenges outlined above, often achieving disease prediction
for efficiently scrutinizing medical imaging data such as accuracy comparable to occasionally surpassing, that of
chest X-rays and computed tomography (CT) scans for conventional radiologists. Among these approaches, CNNs
detecting instances of pneumonia. These models possess have garnered significant attention due to their remarkable
the ability to learn intricate patterns and features from capabilities in image segmentation and classification tasks.
vast datasets to identify abnormal lung conditions infected Leveraging deep learning and computer vision techniques for
with pneumonia with remarkable precision [5]. Further the detection of biomedical issues, particularly in images, has
research and advancements in deep learning are continually proven immensely advantageous. These techniques facilitate
refining pneumonia detection methods, paving the way for swift and accurate disease diagnoses, rivaling the expertise of
more advanced and reliable diagnostic tools in the field of dependable radiologists [10]. In Figure 1, a graphical abstract
respiratory medicine. A significant proportion of pneumonia diagram of pneumonia detection is given, showing various
cases are prevalent in economically disadvantaged and modules of the pneumonia detection system. In this regard,
developing countries, characterized by a dearth of medical this study proposes a CNN model for pneumonia detection
facilities, a dense population, pollution, and unsanitary and makes the following contributions:
surroundings. As a result, averting the potential fatality of • Proposed a deep learning model EfficientNetV2L for
this disease hinges on the timely identification and provision pneumonia detection involving the chest X-ray images
of care. The assessment of respiratory ailments typically from patients. In addition, several state-of-the-art CNN
encompasses radiographic assessment of the lungs, which architectures such as InceptionResNetV2, Xception,
entails methods such as CT, magnetic resonance imaging VGG16 and ResNet50 are utilized for pneumonia
(MRI) and radiography (X-rays). detection using chest X-ray images.
Employing X-ray images presents an affordable and • Experiments involve using a large dataset of 5856 chest
non-intrusive approach to scrutinizing lung conditions [6]. X-ray images with pneumonia and normal classes from a
The interpretation of chest X-rays to detect pneumonia can rich variety of patients from different age groups. Train-
often be ambiguous, leading to potential confusion with test split involves dividing the dataset into training and
other medical conditions. The congestive heart failure and testing sets, with a further 15% split for validation within
lung scarring can imitate pneumonia symptoms. A focused the training set.
assessment of chest X-rays for pneumonia cases might • The study’s focus is on training and evaluating deep
yield misleading results. This inherent challenge contributes learning models on chest X-ray images, aiming to
significantly to the misclassification of X-ray images within improve diagnostic efficiency and enhance patient out-
datasets. Consequently, developing an algorithm capable of comes in clinical settings. The overall goal is to achieve
accurately identifying thoracic disorders, including pneu- accurate pneumonia detection through the proposed
monia holds substantial value in enhancing healthcare deep learning models.
accessibility, especially in remote and underserved regions. • To increase generalizability and robustness of the
The complexity is addressed by evaluating the performance EfficientNetV2L model, k-fold cross-validation is
of diverse pre-trained CNN model versions for the accurate also carried out. Moreover, performance of the
diagnosis of both pathological and normal chest X-rays, EfficientNetV2L model is judged in comparison

34692 VOLUME 12, 2024

M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

FIGURE 1. Graphical abstract of pneumonia detection system.

to existing state-of-the-art models for pneumonia to generalize across diverse cases, ultimately improving its
detection. diagnostic capability. Conventional deep learning models
Pneumonia detection methods, including traditional radio- often lack interpretability, making it challenging for clini-
graphy and even some deep learning models, struggle with cians to trust and understand the decision-making process.
achieving high sensitivity and specificity simultaneously. The EfficientNetV2L model integrates features designed to
False positives and false negatives are common, leading enhance interpretability, facilitating better communication
to both overdiagnosis and underdiagnosis. To proposed between the model and healthcare professionals. Despite this
EfficientNetV2L model aims to address these issues by improvement, there remains a broader need for research into
leveraging its efficient architecture to enhance sensitivity interpretable deep learning models for pneumonia detection.
and specificity, thereby minimizing diagnostic errors. The Some existing deep learning models, especially those with
effectiveness of many existing pneumonia detection methods large architectures, may require substantial computational
is heavily reliant on the quality of the input images. resources and time for training. This limitation hinders
Poor resolution, artifacts, or improper positioning during the practicality of deploying these models in real-time
image acquisition can significantly impact the accuracy of clinical settings. EfficientNetV2L model is designed to
diagnosis. The model, while not completely immune to image balance computational efficiency with performance, making
quality issues, incorporates robust features from Efficient- it more feasible for integration into clinical workflows
NetV2L that contribute to improved performance even in without compromising accuracy. EfficientNetV2L model for
the presence of suboptimal image conditions. This limitation pneumonia detection but also underscores the limitations of
by utilizing a well-curated dataset and leveraging the current pneumonia detection methods. By acknowledging
efficiency of EfficientNetV2L to enhance the model’s ability these challenges, this research aims to contribute to the

VOLUME 12, 2024 34693

 M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

ongoing efforts to advance the field of medical imaging appropriate treatment for the specific organism causing the
and improve the accuracy and efficiency of pneumonia infection [15].
diagnosis.
The rest of the paper is distributed into the following B. ROLE OF MACHINE LEARNING IN PNEUMONIA
sections: Section II discusses the related work that presents CLASSIFICATION
the literature of the current research works and technologies.
Using chest X-ray images for pneumonia classification,
Section III presents the methodology that involves the
machine learning in particular deep learning proves to be very
overall research approach, describing the data collection
promising. Deep learning models such as CNN-based models
methods and discussing the data analysis techniques utilized.
have demonstrated a superior ability to learn intricate patterns
Section IV shows the results and discussion of the proposed
and features from large datasets. In the context of pneumonia
approach. Section V presents the supporting points and
detection, these models are trained on diverse chest X-ray
establishes a connection to the broader context.
images, enabling them to identify specific visual cues and
abnormalities associated with pneumonia. By analyzing the
II. RELATED WORK
unique characteristics of pneumonia cases, deep learning
Pneumonia is a common respiratory illness characterized by
algorithms can accurately classify chest X-ray images
lung inflammation and infection, often a condition brought
as pneumonia or non-pneumonia cases. This automated
on by pathogens like bacteria, viruses or fungus. For the
classification process facilitates prompt identification of
disease to be effectively treated and managed, pneumonia
potential pneumonia cases, aiding in early intervention and
must be detected as early and accurately as possible. Deep
timely treatment [31]. Machine learning improves pneumonia
learning techniques have recently emerged as a promising
detection efficiency, accuracy and patient outcomes through
approach for automating pneumonia detection using chest
advancements in deep learning algorithms and annotated
X-ray images. By leveraging CNNs and training them on
datasets, leading to more effective diagnoses and improved
diverse datasets, deep learning algorithms can learn complex
healthcare practices [32]. Pneumonia is a deadly infec-
patterns and features indicative of pneumonia. This enables
tion affecting young and elderly individuals. Researchers
them to identify specific visual markers and abnormalities
identified it using X-ray scans, algorithms and traditional
associated with the disease [11]. Deep learning integration
machine learning classifiers. A mask-RCNN-based model
in pneumonia detection can improve diagnostic accuracy,
demonstrated reliability and potency [33]. Through the use
minimize human error and improve healthcare efficiency.
of machine learning such as support vector machine (SVM),
Research and annotated datasets will refine deep learning
K nearest neighbor (KNN), decision tree (DT) and deep
models, enabling more reliable and effective detection from
learning such as multilayer perceptron (MLP) and RNN
chest X-ray images. This approach considers symptoms
models, the authors studied the prediction of pneumonia
severity and potential complications, allowing radiologists to
disease and obtained promising accuracy.
identify infiltrates, white patches in lung tissue and potential
pneumonia-related complications like abscesses or pleural
effusions [12]. C. ROLE OF DEEP LEARNING IN PNEUMONIA DISEASES
CLASSIFICATION
A. X-RAY IMAGES USED IN PNEUMONIA DISEASES The categorization of pneumonia infections using chest X-ray
Deep learning models are essential for detecting and diag- pictures depends heavily on deep learning. Its utilization,
nosing pneumonia using chest X-ray images. These images particularly through CNN, has revolutionized pneumonia
provide visual insights into the lung’s condition, allowing detection. Deep learning algorithms excel at learning
algorithms to recognize patterns associated with pneumonia. intricate patterns and features from extensive datasets,
Training deep learning models on diverse datasets helps enabling them to extract meaningful information from
distinguish between normal lung conditions and those chest X-ray images [34]. The integration of deep learning
indicative of pneumonia [13]. The information-rich nature in pneumonia disease classification empowers healthcare
of X-ray images, combined with the power of deep learning professionals with an automated and efficient tool for early
techniques, empowers healthcare practitioners with an auto- diagnosis [35].
mated and efficient tool for accurate pneumonia detection. In the present context, the classification of medical images
This technology facilitates early intervention, enabling timely has emerged as a particularly captivating and inspiring
treatment and improved patient outcomes [14]. Pneumonia endeavor. Leveraging deep learning methodologies, a range
diagnosis involves various techniques, including CT scans, of six distinct diseases, including one non-tissue infection-
sputum tests, pulse oximetry, thoracentesis, blood gas related ailment, are detected and classified. Notably, neural
analysis, bronchoscopy, pleural fluid cultures and complete networks are specifically employed for the identification
blood counts. Treatment strategies are tailored to the and categorization of interstitial lung diseases (ILDs) [36].
underlying cause, with antibiotics for bacterial pneumonia, Another article describes the Boltzmann machine, a device
antiviral medications for viral pneumonia and antifungal that uses lung-generated tomography (CT) to analyze
medications for fungal pneumonia. This approach ensures illnesses. Texture classification and airway detection are

34694 VOLUME 12, 2024

M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

TABLE 1. Performance comparison with existing stateoftheart models.

two novel strategies that are introduced in the study using deep CNN with dense connections and a 3D deep CNN
two datasets. CNN networks are employed to build on the with shortcut connections. The gradient vanishing issue is
kernel to evaluate the patient’s brain tissue. They use a dual resolved by shortcuts and dense connections. Deep networks
framework, for example using CNN to feature extraction can capture both the general and specific characteristics of
first and then classifying images based on the feature data. lung nodules because of connections [42]. A CNN-based
In a different study, the scientists created several models to approach combined with rib suppression and lung field
support the accuracy of the top model’s pneumonia detection segmentation is applied. For pixel patches obtained in the
results [37]. They used a dataset of 26,684 images to train lung area, three CNN models are trained on images of
the CNN models AlexNet, LeNet, GoogleNet, ResNet and varied resolutions, and feature fusion is used to merge all
VGGNet using a 1024×1024 image resolution. Their study’s the data [43]. Image augmentation and a region-based CNN
findings showed that the VGGNet model had an accuracy rate are employed to partition the pulmonary pictures and identify
of 97%, while the ResNet model had the lowest accuracy rate pneumonia in [44]. Image segmentation is the division of
at 74% [38]. Similarly, the authors employed mask R-CNN an image into a number of smaller sections. Typically, the
and RetinaNet, two different types of CNN, and got good properties of the pixels in these sub-regions are similar and
results [39]. they do not cross [45]. With the ongoing development of
modern computer science and medical technology, medical
D. SEGMENTATION-BASED PNEUMONIA DETECTION images are now divided into separate image types that are
Segmentation-based pneumonia detection using deep learn- different from standard images. This is an important area
ing has emerged as a promising methodology in the field of application for picture segmentation [46]. These elements
of medical imaging. This approach focuses on accurately lead to interesting tissues appearing in some segments.
identifying and segmenting the regions of pneumonia within The contrast lessens and the edge part becomes rather
chest X-ray images. By leveraging advanced deep learning ambiguous and uneven. Therefore, accurate medical image
architectures such as U-Net or Mask R-CNN, these models segmentation is a very difficult scientific research task.
can effectively learn to segment and isolate the areas of Medical image segmentation typically calls for extremely
the lungs affected by pneumonia. This segmentation process high results regarding integrity and accuracy [47]. Following
provides crucial insights into the location and extent of the the completion of the preliminary imaging, specific technical
infection, facilitating precise diagnosis and treatment plan- tools are typically required to process the results appropri-
ning [40]. Segmentation-based pneumonia detection using ately in order to show distinct and recognizable tissues of
deep learning allows healthcare professionals to visualize and interest [48]. The comparative related work is described in
analyze specific regions, enabling more accurate and targeted Table 15.
interventions. This approach quantifies pneumonia severity
and progression, enabling efficient monitoring and follow- III. MATERIALS AND METHODS
up care. Research and advancements in segmentation-based The methodology for pneumonia detection using deep
pneumonia detection hold promise for improving diag- learning based on chest X-ray images involves several key
nostic capabilities and patient outcomes in respiratory components, as shown in Figure 2. Six deep learning mod-
medicine. els, such as CNN, InceptionResNetV2, Xception, VGG16,
Mask-RCNN is a deep neural network used for pulmonary ResNet50 and EfficientNetV2L have been used in this
image segmentation, image augmentation, dropout and L2 study. These models are implemented and fine-tuned using
regularization detecting pneumonia using both global and the Tensor Flow library, a widely adopted deep learning
local characteristics [41]. The authors used both a 3D framework. In order to improve the model’s robustness and

VOLUME 12, 2024 34695

 M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

FIGURE 2. Methodology of pneumonia detection using efficient NetV2L model.

FIGURE 3. Chest x-ray images for pneumonia detection.

prevent overfitting, data augmentation techniques including X-ray imaging was performed as part of the patients’ routine
random rotations, shifts, flips and zooms to the original clinical care. For the analysis of chest X-ray images, all chest
images have been applied to generate additional training radiographs were initially screened for quality control by
samples. Furthermore, the X-ray images underwent pre- removing all low-quality or unreadable scans. The diagnoses
processing steps such as normalization of pixel values and for the images were then graded by two expert physicians
resizing to a standardized input size. This comprehensive before being cleared for training in the AI system. In order
methodology allowed us to leverage diverse deep learning to account for any grading errors, the evaluation set was
models, implement effective data augmentation, preprocess also checked by a third expert [49]. This study leverages
the chest X-ray images, efficiently train and evaluate the several models to perform accurate pneumonia detection,
pneumonia detection models. improving diagnostic efficiency and potentially enhancing
patient outcomes. A total of 5,856 samples, including 1,583
A. DATASET DESCRIPTION normal and 4,273 pneumonia images, were used in this study.
The dataset contains 5,863 X-ray images (JPEG) and The chest X-ray scans of the volunteer patients are used to
2 categories (pneumonia and normal). Chest X-ray images create the dataset. The rationale for using the dataset is its
(anterior-posterior) were selected from retrospective cohorts usage in several existing state-of-the-art approaches. For a
of pediatric patients of one to five years old from Guangzhou fair comparison with these methods, the dataset can be used
Women and Children’s Medical Center, Guangzhou. All chest as a benchmark. In addition, the dataset contains a rich variety

34696 VOLUME 12, 2024

M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

of X-ray images from patients and is well-suited to analyze ability to detect pneumonia accurately. By applying trans-
the performance of the proposed model. Table 2 shows the formations such as rescale, rotation_range, horizontal_flip,
number of images used for training, testing and validation. vertical_flip, height_shift_range and width_shift_range. The
dataset’s diversity is enriched, aiding the model in learning
TABLE 2. Dataset details for pneumonia and normal images. a more comprehensive set of features. Additionally, resizing
images is a crucial preprocessing step, ensuring uniformity in
input dimensions for the neural network. Image augmentation
serves as a valuable tool in pneumonia detection, enhancing
the model’s capability to capture essential features and
Figure 3 shows X-ray images of the patients. Each ultimately improving the accuracy and reliability of the deep
panel corresponds to an individual patient and displays learning-based system.
their respective X-ray images of normal or pneumonia
radiographs. B. DEEP LEARNING MODELS
In deep learning models, the landscape of pneumonia
1) IMAGE PREPROCESSING detection has undergone a revolutionary transformation by
Image preprocessing plays a critical role in detecting learning intricate patterns and features inherent in chest
pneumonia using deep learning with chest X-ray images. X-ray images. CNNs have emerged as frontrunners, largely
It involves a series of essential operations applied to the owing to their remarkable proficiency in image classification
images prior to inputting them into the deep learning models. endeavors. The realm of pneumonia detection has wit-
The goal is to improve the quality of images and the nessed the successful deployment of advanced architectures,
consistency of the data. Typical techniques include resizing such as InceptionResNetV2, Xception, VGG16, ResNet50
the images to a standardized resolution, normalizing pixel and EfficientNetV2L. These models incorporate ingenious
values to a specific range, and applying contrast adjustments techniques like residual connections, depth-wise separable
to improve image clarity. Additionally, noise reduction convolutions, and efficient scaling to elevate the accuracy
methods like Gaussian or median filtering can be employed to of feature extraction. The amalgamation of these deep
eliminate unwanted artifacts. Other methods of increasing the learning models into pneumonia detection holds the promise
dataset and enhancing model generalization include rotation, of substantially expediting diagnoses, improving treatment
scaling and flipping of the images. By preprocessing the outcomes and enhancing healthcare practices in the domain
images, the input data becomes more suitable for deep of respiratory medicine. The selection of CNN models
learning algorithms, enabling better feature extraction and is motivated by the results reported in existing studies
improving the model’s ability to identify pneumonia-related concerning the use of these models on image datasets.
patterns. Effective image preprocessing is crucial in optimiz- For example, [50] reports better results using a custom
ing the performance and accuracy of pneumonia detection CNN model for X-ray-based pneumonia detection. The
models, ultimately leading to more reliable diagnoses and authors advocate the use of GoogleNet, ResNet-18 and
enhanced patient care. DenseNet-121 for pneumonia detection in [6]. Similarly,
AlexNet, ResNet18, DenseNet201 and SqueezeNet are used
2) IMAGE AUGMENTATION for the same task in [40] showing better performance than
Image augmentation plays an important role in pneumonia other deep learning models. The study [51] presents the
detection using deep learning with chest X-ray images. performance analysis of deep learning models in the context
It involves applying various transformations to the original of pneumonia detection and suggests using CNN variants for
images to expand the training dataset and enhance the better performance. The current dataset is selected to make
model’s performance. By introducing variations in the data, a fair comparison with existing state-of-the-art methods for
image augmentation helps improve the model’s ability to pneumonia detection. The dataset serves as the benchmark,
generalize to unseen images. Leveraging tools such as as it has been utilized by existing studies.
ImageDataGenerator, a versatile module in deep learning Pneumonia detection using CNN models plays an indis-
frameworks like TensorFlow and Keras, allows for the pensable role, particularly in the analysis of chest X-ray
generation of augmented image data on the fly during model images. These models excel at extracting pertinent features
training. Common augmentation techniques include rotation, from input images, subsequently pooling these features
translation, scaling, flipping and adding random noise. These to facilitate classification tasks. The inherent depth of
transformations simulate real-world variations in chest X-ray their architecture empowers them to acquire hierarchical
images, accounting for different orientations, positions and representations of chest X-ray images, thereby capturing
noise levels. Additionally, image augmentation helps address intricate pneumonia-related patterns. By training on extensive
the issue of imbalanced datasets where the number of datasets containing annotated images, diagnostic precision
pneumonia cases may be limited compared to normal cases. and efficiency are significantly elevated, enabling healthcare
By generating additional samples through augmentation, the practitioners to make well-timed and informed decisions. The
model can learn from a more balanced dataset, improving its continuous evolution of CNN architectures, along with the

VOLUME 12, 2024 34697

 M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

FIGURE 4. EfficientNetV2L architectural diagram.

availability of annotated datasets, continues to augment the discernible patterns and irregularities associated with pneu-
effectiveness of pneumonia detection through deep learning. monia. Their remarkable proficiency in accurately classifying
This progress not only contributes to enhanced patient X-ray images into either pneumonia or non-pneumonia
outcomes but also fosters more streamlined and effective categories has significantly bolstered the accuracy of diag-
healthcare practices. noses and instilled greater reliability. Looking ahead, the
Leveraging InceptionResNetV2 models proves to be a ongoing progress and evolution of VGG16 models hold the
formidable approach to pneumonia detection through deep potential to further refine pneumonia detection, ultimately
learning, especially concerning chest X-ray images. The translating into improved patient outcomes and more stream-
amalgamation of Inception and ResNet attributes empowers lined healthcare practices within the realm of respiratory
these models to meticulously extract intricate information medicine.
from X-ray images by employing convolutional layers and ResNet50 models for pneumonia detection through deep
residual connections. The outcome is a precise classifica- learning using chest X-ray images prove to be a potent
tion of X-ray images into pneumonia or non-pneumonia approach. Their architecture, characterized by the incorpora-
categories, thus elevating the accuracy of diagnoses and tion of residual connections, empowers them to grasp com-
offering indispensable support to healthcare experts. As the plex patterns and intricate features. With extensive dataset
capabilities of InceptionResNetV2 models continue to training, ResNet50 models adeptly recognize discernible
evolve, they harbor significant potential for further refining visual cues that signify the presence of pneumonia. The
pneumonia detection within the realm of deep learning. This outcome is an improvement in the accuracy and reliability
trajectory holds the promise of delivering superior patient of diagnoses, offering invaluable support to healthcare
outcomes and fostering more streamlined and effective experts in their endeavor to identify and address pneumonia
healthcare practices, particularly in the domain of respiratory cases based on chest X-ray images. As ResNet50 models
medicine. continue to advance and evolve, their potential for further
Xception models for pneumonia detection through deep enhancing pneumonia detection remains promising. This
learning with chest X-ray images are a potent strategy. trajectory holds the potential for elevating patient outcomes
These models boast an architecture that employs depth-wise and fostering more effective healthcare practices, particularly
separable convolutions, facilitating the nuanced capture of within the domain of respiratory medicine.
intricate patterns and distinctive features. Through harnessing
substantial datasets, Xception models adeptly discern visual C. PROPOSED EFFICIENTNETV2L
attributes closely tied to pneumonia. This proficiency is EfficientNetV2L models for pneumonia detection using chest
reflected in their remarkable precision in precisely catego- X-ray images prove to be an effective strategy. These models
rizing chest X-ray images into pneumonia or non-pneumonia seamlessly integrate scaling techniques and deep neural
cases, thereby elevating the accuracy of diagnoses. As Xcep- networks, yielding a synergy that results in both efficiency
tion models continue to evolve and advance, their potential and accuracy. With robust training on extensive datasets
for refining diagnostic capabilities and ultimately enhancing containing annotated chest X-ray images, EfficientNetV2L
patient outcomes in the realm of respiratory medicine holds models adeptly capture the distinct visual characteristics
great promise. associated with pneumonia. This aptitude translates into
VGG16 models stand as potent tools in the domain an enhancement in diagnostic efficiency and reliability,
of pneumonia detection through deep learning with chest providing invaluable assistance to healthcare practitioners in
X-ray images. Their architecture, characterized by a series identifying and addressing pneumonia cases through chest
of convolutional layers and max-pooling layers, excels at X-ray images. As the advancement of EfficientNetV2L
extracting intricate hierarchical features. With the advantage models continues, their potential to further refine pneumonia
of extensive dataset training, VGG16 models adeptly discern detection holds immense promise. This trajectory not only

34698 VOLUME 12, 2024

M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

(TP)
bears the potential to elevate patient outcomes but also to Recall = (3)
foster more streamlined and effective healthcare practices (TP + FP)
within the realm of respiratory medicine. Figure 4 shows 2 ∗ (Precision ∗ Recall)
F1 score = (4)
the architecture of the proposed EfficientNetV2I model for (Precision + Recall)
pneumonia detection.
Architectural details of the proposed model are given in IV. RESULTS AND DISCUSSION
Table 3 along with the details of input, functional, global, The performance of deep learning models hinges on their
and dense layers. It also includes information on output ability to accurately process and interpret data to generate
shape, parameters for each layer and hyperparameters. The predictions. To gauge this performance, diverse metrics come
EfficientNetV2L architecture was fine-tuned for pneumonia into play, encompassing accuracy, precision, recall, and
detection from chest X-rays. Leveraging the power of deep the F1 score. Accuracy serves as a barometer of overall
learning, the potential of EfficientNetV2L, known for its correctness in predictions. Precision quantifies the ratio of
efficiency and performance in image classification tasks, was true positive cases to all predicted positives, while recall
recognized. The approach involved training the model on a emphasizes the reduction of false positives. These metrics
diverse and extensive dataset of chest X-rays, meticulously collectively provide a comprehensive assessment of a model’s
labeled for pneumonia cases. efficacy in various aspects of data processing and prediction.

TABLE 3. EfficientNetV2L model configuration parameters used in the A. RESULTS OF CONVOLUTIONAL NEURAL NETWORK
model training process along with their respective values and
descriptions. CNN exhibits remarkable efficacy in tackling classification
tasks, with their evaluation encompassing metrics such as
accuracy, precision, recall and the F1 score. These networks
learn patterns and features from chest X-ray datasets and
undergo assessment using distinct test datasets. Accuracy
quantifies the ratio of correctly classified instances to the
overall sample size, while precision computes the fraction
of accurately predicted positive instances. Through the opti-
mization of weights and biases, CNNs achieve elevated levels
The fine-tuning process aimed to optimize the net-
of accuracy, precision, recall and F1 scores. Renowned for
work’s parameters to enhance its sensitivity and specificity
their proficiency, CNNs serve as invaluable tools in diverse
in identifying pneumonia-related patterns in radiographic
classification tasks, particularly in scenarios involving image
images. Table 4 showed the EfficientNetV2L fine-tuned for
data.
pneumonia detection from chest X-rays. Table 5 show the
The CNN algorithm demonstrated strong performance,
comaprison of EfficientNetV2L with other models in terms
achieving an accuracy of 88.78%, precision of 91.45%,
of working and performance.
recall of 90.51% and F1 score of 90.97%. These results
indicate the model’s ability to accurately predict outcomes
D. EVALUATION PARAMETERS
and effectively capture positive instances as shown in Table 6.
The assessment of pneumonia detection models utilizing
In addition, Figure 5 shows the confusion matrix for the CNN
deep learning with chest X-ray images encompasses the
model. It indicates that the model has 554 correct predictions
evaluation of various metrics, including accuracy, recall,
including 201 true positives and 353 true negatives while the
precision and the F1 score. Accuracy gauges the model’s
model made 70 wrong predictions.
proficiency in accurately predicting positive cases, while
specificity evaluates its aptitude in correctly identifying
negative cases. Precision quantifies the fraction of accurately
classified positive cases out of all predicted positive cases.
The F1 score presents a harmonized measure by balancing
both precision and recall, thereby offering healthcare prac-
titioners a comprehensive tool to inform their decisions and
ultimately enhance patient care. The following equations are
used for accuracy, precision, recall and F1 score:
(TP + TN )
Accuracy = (1)
(TP + TN + FP + FN )
The evaluation of the model’s accuracy involves the
examination of true positives (TP), true negatives (TN), false
positives (FP) and false negatives (FN).
(TP)
Precision = (2) FIGURE 5. Confusion matrix for CNN model.
(TP + FP)
VOLUME 12, 2024 34699
 M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

TABLE 4. Fine-tuning of the EfficientNetV2L for pneumonia detection from chest X-rays.

TABLE 5. Comparison of EfficientNetV2L model with other models.

TABLE 6. Results of CNN performance. its robust feature extraction capabilities, elevated accuracy
rates and sophisticated design, this model guarantees precise
predictions, impressive precision scores and outstanding
F1 scores. Consequently, InceptionResNetV2 emerges as a
dependable and commendable choice for classification tasks
centered on image data.
The InceptionResNetV2 model demonstrated strong per-
formance, achieving an accuracy of 88.94%, precision of
B. RESULTS OF INCEPTIONRESNETV2 MODEL 89.92%, recall of 88.94% and F1 score of 88.55%. These
InceptionResNetV2, a formidable deep learning model, results indicate the model’s ability to accurately predict
seamlessly integrates the architectural strengths of Inception outcomes and effectively capture positive instances as shown
and ResNet paradigms to facilitate precise evaluations in Table 7. In addition, Figure 6 shows the number of correct
encompassing accuracy, precision and the F1 score. With and wrong predictions from the InceptionResNetV2 model.

34700 VOLUME 12, 2024

M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

TABLE 7. Results of InceptionResNetV2 model for pneumonia detection. performance of the Xception model, compared to both CNN
and InceptionResNetV2 model with 566 correct predictions
while 58 predictions are wrong.

The model has comparatively 69 wrong predictions while

the correct predictions are 555. The number of true negative
samples is higher compared to the CNN model.

FIGURE 7. Confusion matrix for Xception model.

D. RESULTS FOR VGG16 MODEL

VGG16, a deep learning model comprising layers distin-
guishes itself through its exceptional aptitude in image
classification tasks. The model adeptly captures complex
patterns and features, leading to the attainment of elevated
FIGURE 6. Confusion matrix for InceptionResNetV2 model. accuracy rates and precision scores. A defining characteristic
of VGG16 is its capacity to achieve a harmonious equilibrium
between precision and recall, culminating in the generation of
C. RESULTS FOR XCEPTION MODEL impressive F1 scores.
Xception stands as a robust and influential deep learning
model tailored for image classification tasks, renowned for TABLE 9. Performance results for VGG16 model.
its remarkable prowess in accuracy, precision and F1 score
evaluations. Distinguished by its novel ‘‘depth-wise separable
convolution’’ approach, Xception adeptly captures both spa-
tial and channel-wise dependencies, facilitating the extraction
of meaningful features and ultimately attaining elevated
accuracy rates. The model excels in striking a harmonious
equilibrium between precision and recall, rendering it a The VGG16 model demonstrated strong performance on
dependable and trustworthy option for classification tasks the testing set, achieving an accuracy of 91.66%, precision
centered on image data. of 92.08%, recall of 91.66% and F1 score of 91.49%. These
results indicate the model’s ability to accurately predict
TABLE 8. Results of Xception model performance. outcomes and effectively capture positive instances as shown
in Table 9.
Figure 8 shows the network’s confusion matrices of the
VGG16 model. It shows that there is a higher difference
between the performance of the VGG16 and the Xception
model. VGG16 has 52 wrong predictions compared to
58 wrong predictions from the Xception model. Similarly, the
The Xception model demonstrated strong performance on number of correct predictions, i.e., 572, is also high using the
the testing set, achieving an accuracy of 90.70%, precision VGG16 model.
of 90.72%, recall of 90.70% and F1 score of 90.61%. These
results indicate the model’s ability to accurately predict E. RESULTS FOR RESNET50 MODEL
outcomes and effectively capture positive instances as shown ResNet50 stands as a robust deep learning model distin-
in Table 8. In addition, Figure 7 shows the confusion matrix of guished for its remarkable prowess in accuracy, precision,
the Xception model. The confusion matrix indicates a better recall and F1 score assessments within image classification

VOLUME 12, 2024 34701

 M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

TABLE 11. Performance of EfficientNetV2L model.

TABLE 12. Analysis of all employed models.

FIGURE 8. Confusion matrix for VGG16 model.

TABLE 13. Results of deep learning models using hyperparameter tuning.

tasks. Characterized by its intricate layer architecture,

the model adeptly captures complex patterns and features
crucial for accurate classifications. Unique to ResNet50 are
its skip connections, which effectively mitigate vanishing
gradient challenges, enhancing training stability. The model
effectively harmonizes precision and recall, culminating in
the generation of impressive F1 scores. TABLE 14. Model performance with k = 10.

TABLE 10. Performance of ResNet50 for pneumonia detection.

ResNet50 model. Results show that the performance of

ResNet50 is poorer than those of CNN, InceptionResNetV2,
Xception and VGG16 with 75 wrong predictions.

F. RESULTS OF PROPOSED EFFICIENTNETV2L MODEL

EfficientNetV2L is a widely esteemed deep learning model
celebrated for its outstanding proficiency in accuracy, preci-
sion, recall and F1 score assessments within the domain of
image classification tasks. Distinguished by its streamlined
architecture, compound scaling techniques, integration of
neural architecture search, and strategic utilization of building
blocks, the model excels in achieving elevated accuracy rates.
It not only prioritizes the generation of precise predictions but
FIGURE 9. Confusion matrix for ResNet50 model. also adeptly harmonizes precision and recall, culminating in
the production of impressive F1 scores.
The ResNet50 model demonstrated strong performance on The EfficientNetV2L model demonstrated strong per-
the testing set, achieving an accuracy of 87.98%, precision formance on the testing set, achieving an accuracy of
of 88.14%, recall of 87.98% and F1 score of 87.75%. These 94.02%, precision of 94.40%, recall of 97.24% and F1 score
results indicate the model’s ability to accurately predict of 95.80%. These results indicate the model’s ability to
outcomes and effectively capture positive instances as shown accurately predict outcomes and effectively capture positive
in Table 10. Figure 9 shows the confusion matrices of the instances as shown in Table 11.

34702 VOLUME 12, 2024

M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

FIGURE 10. Performance analysis of deep learning models.

FIGURE 11. Performance regarding ROC-AUC curve.

G. COMPARATIVE ANALYSIS OF ALL MODELS performed exceptionally well with impressive performance
Table 12 shows the outcomes of all models used in this study. evaluation metrics.
Results are provided for accuracy, precision, recall and F1
scores. The suggested model outperforms every other model, H. EVALUATION USING HYPERPARAMETER TUNING
including CNN, InceptionResNetV2, Xception, VGG16, Hyperparameters in deep learning are variables that are either
ResNet50 and EfficientNetV2L, which are all fine-tuned to manually set by the user or chosen via hyperparameter
get the best performance with the available dataset. However, optimization methods. A deep learning model’s architecture
compared to all other models in use, the performance of the and optimization are determined by these parameters. Hyper-
proposed model is superior. parameters have an impact on the learning process itself,
Figure 10 shows a visual representation of how well each as opposed to model parameters, which are learned during
deep learning model performed for pneumonia detection. the training phase. Table 13 provides the best performance of
It can be observed that the proposed EfficientNetV2L model each model as obtained using hyperparameter tuning.

VOLUME 12, 2024 34703

 M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

TABLE 15. Performance comparison of EfficientNetV2L model with existing state-of-the-art models.

TABLE 16. Compare analysis of different technologies with the proposed approach.

I. ROC-AUC OF MODELS PERFORMANCE across different folds, however, its average performance is
To purify insights and render prudent judgments, proposed better than other models.
data-driven solutions leverage an array of neural network
architectures, encompassing CNN, InceptionResNetV2, K. COMPARATIVE ANALYSIS WITH EXISTING MODELS
Xception, VGG16, ResNet50 and EfficientNetV2L. Eval- Comparative analysis is carried out between various models
uation of model efficacy is conducted through ROC-AUC from existing literature that performed pneumonia detection
and PR-AUC metrics, ensuring resilient classification using deep learning models. These studies are selected
performance and a thorough grasp of model dynamics, because they use the same dataset for experiments and can
exemplified in Figures 11. The precision parameter delineates be used as the benchmark to evaluate the performance of
the accurately predicted values vis-à -vis the total predictions, the proposed EfficientNetV2L model. Table 15 highlights the
elucidating the discernment of identified attacks in real- performance variations of different deep learning models and
world systems, elucidated in the result section tables. approves the superior performance of the EfficientNetV2L
EfficientNetV2L, distinguished by its exceptional scalability model.
and efficiency, notably excels in attaining a commendable
ROC-AUC, underscoring its proficiency in addressing L. COMPARE ANALYSIS OF DIFFERENT STUDIES WITH
intricate classification challenges with optimal efficacy. PROPOSED APPROACH
Although several approaches and technologies have been
J. RESULTS OF K-FOLD CROSS-VALIDATION investigated in this study, as shown in Table 16. The proposed
For performance validation, k-fold cross-validation is carried approach integrates the finest features to provide a thorough
out for the proposed approach. Table 14 shows the results with and effective solution to the issue at hand. To develop a
the training time and for k = 10, respectively. Results indicate cutting-edge strategy that has the potential to dramatically
that the performance of the proposed models varies slightly improve the area by combining insights and methods from

34704 VOLUME 12, 2024

M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

these many sources. The contribution of the proposed method [3] Y. Muhammad, M. D. Alshehri, W. M. Alenazy, T. V. Hoang, and
demonstrates the value of multidisciplinary cooperation and R. Alturki, ‘‘Identification of pneumonia disease applying an intelligent
computational framework based on deep learning and machine learning
innovative problem-solving, paving the way for promising techniques,’’ Mobile Inf. Syst., vol. 2021, pp. 1–20, May 2021.
future advances. [4] N. M. Elshennawy and D. M. Ibrahim, ‘‘Deep-pneumonia framework using
deep learning models based on chest X-ray images,’’ Diagnostics, vol. 10,
no. 9, p. 649, Aug. 2020.
V. CONCLUSION
[5] Standardization of Interpretation of Chest Radiographs for the Diag-
In conclusion, the classification of digital chest X-ray images nosis of Pneumonia in Children, World Health Organization, Geneva,
using deep learning techniques for pneumonia detection is Switzerland, 2001.
proposed. The research implemented several CNN models, [6] R. Kundu, R. Das, Z. W. Geem, G.-T. Han, and R. Sarkar, ‘‘Pneumonia
detection in chest X-ray images using an ensemble of deep learning
including CNN InceptionResNetV2, Xception, VGG16, models,’’ PLoS ONE, vol. 16, no. 9, Sep. 2021, Art. no. e0256630.
ResNet50 and EfficientNetV2L using Python programming [7] D. Varshni, K. Thakral, L. Agarwal, R. Nijhawan, and A. Mittal,
and using the tools Google Colab. Initial experiments showed ‘‘Pneumonia detection using CNN based feature extraction,’’ in Proc.
IEEE Int. Conf. Electr., Comput. Commun. Technol. (ICECCT), Feb. 2019,
promising results, with the models achieving accuracy rates pp. 1–7.
of 88.78%, 88.94%, 90.7%, 91.66%, 87.98% and 94.02%. [8] M. Grewal, M. M. Srivastava, P. Kumar, and S. Varadarajan, ‘‘RADnet:
These accuracy values indicate the potential of these models Radiologist level accuracy using deep learning for hemorrhage detection in
to serve as decision support tools for pneumonia diagnosis CT scans,’’ in Proc. IEEE 15th Int. Symp. Biomed. Imag. (ISBI), Apr. 2018,
pp. 281–284.
based on X-ray images. However, the paper acknowledges [9] A. Hosny, C. Parmar, J. Quackenbush, L. H. Schwartz, and H. J. Aerts,
some limitations and highlights the need for further research. ‘‘Artificial intelligence in radiology,’’ Nat. Rev. Cancer, vol. 18, no. 8,
One concern is the possibility of overfitting due to the pp. 500–510, 2018.
dataset’s size. To address this, the authors suggest gathering [10] S. Naicker, J. Plange-Rhule, R. C. Tutt, and J. B. Eastwood, ‘‘Shortage of
healthcare workers in developing countries—Africa,’’ Ethnicity Disease,
more data to enhance the model’s generalization ability. vol. 19, pp. 60–64, Apr. 2009.
The paper also emphasizes the importance of exploring [11] A. M. A. Barhoom and S. S. Abu-Naser, ‘‘Diagnosis of pneumonia using
various preprocessing techniques and CNN configurations to deep learning,’’ Int. J. Acad. Eng. Res., vol. 6, no. 2, pp. 48–68, 2022.
improve the models’ performance. Additionally, data aug- [12] A. M. A. Barhoom, A. J. Khali, B. S. Abu-Nasser, M. N. Musleh, and
S. S. Abu-Naser, ‘‘Predicting Titanic survivors using artificial neural
mentation techniques are proposed to increase the dataset’s network,’’ Int. J. Acad. Eng. Res., vol. 3, no. 9, pp. 8–12, 2019.
size, which can aid in better generalization. Furthermore, the [13] Revised Who Classification and Treatment of Pneumonia in Children
authors suggest incorporating additional X-ray datasets that at Health Facilities: Evidence Summaries, World Health Organization,
Geneva, Switzerland, 2014.
include data labels for various pathologies. This expansion
[14] J. Gao, Z. Tian, and X. Yang, ‘‘Breakthrough: Chloroquine phosphate has
would enable the models to differentiate between differ- shown apparent efficacy in treatment of COVID-19 associated pneumonia
ent lung abnormalities, enhancing accuracy and reliability. in clinical studies,’’ BioScience Trends, vol. 14, no. 1, pp. 72–73, 2020.
By conducting a series of experiments, want to provide the [15] P. Zhou, X. L. Yang, X. G. Wang, B. Hu, L. Zhang, W. Zhang, H. R. Si,
Y. Zhu, B. Li, C. L. Huang, and H. D. Chen, ‘‘A pneumonia outbreak
best pre-trained models and classifiers for use in further associated with a new coronavirus of probable bat origin,’’ Nature, vol. 579,
research in a related field. The EfficientNetV2L model, no. 7798, pp. 270–273, Mar. 2020.
which uses chest radiographs for pneumonia detection, offers [16] M. Aledhari, S. Joji, M. Hefeida, and F. Saeed, ‘‘Optimized CNN-based
potential for future advancement in medical imaging and diagnosis system to detect the pneumonia from chest radiographs,’’ in Proc.
IEEE Int. Conf. Bioinf. Biomed. (BIBM), Nov. 2019, pp. 2405–2412.
artificial intelligence. It can be improved by incorporating [17] A. I. Khan, J. L. Shah, and M. M. Bhat, ‘‘CoroNet: A deep neural
multi-modal data, incorporating clinical information, and network for detection and diagnosis of COVID-19 from chest X-ray
exploring transfer learning from related medical domains. images,’’ Comput. Methods Programs Biomed., vol. 196, Nov. 2020,
Art. no. 105581.
The model’s interpretability and explainability aspects can
[18] M. M. Hasan, M. M. J. Kabir, M. R. Haque, and M. Ahmed, ‘‘A
address trust and acceptance concerns in clinical settings. Its combined approach using image processing and deep learning to detect
scalability across diverse populations and demographics is pneumonia from chest X-ray image,’’ in Proc. 3rd Int. Conf. Electr.,
crucial for its effectiveness. The model’s potential for early Comput. Telecommun. Eng. (ICECTE), Dec. 2019, pp. 89–92.
[19] V. Sirish Kaushik, A. Nayyar, G. Kataria, and R. Jain, ‘‘Pneumonia
detection and prognosis prediction could significantly impact detection using convolutional neural networks (CNNs),’’ in Proc. 1st
patient outcomes. Int. Conf. Comput., Commun., Cyber-Secur. Singapore: Springer, 2020,
pp. 471–483.
[20] M.-J. Tsai and Y.-H. Tao, ‘‘Machine learning based common radiologist-
ACKNOWLEDGMENT
level pneumonia detection on chest X-rays,’’ in Proc. 13th Int. Conf. Signal
(Mudasir Ali, Mobeen Shahroz, and Urooj Akram contributed Process. Commun. Syst. (ICSPCS), 2019, pp. 1–7.
equally to this work.) [21] K. Wang, P. Jiang, J. Meng, and X. Jiang, ‘‘Attention-based DenseNet
for pneumonia classification,’’ IRBM, vol. 43, no. 5, pp. 479–485,
Oct. 2022.
REFERENCES [22] I. U. Khan, S. Azam, S. Montaha, A. Al Mahmud, A. K. M. R. H. Rafid,
[1] V. Chouhan, S. K. Singh, A. Khamparia, D. Gupta, P. Tiwari, C. Moreira, Z. Hasan, and M. Jonkman, ‘‘An effective approach to address processing
R. Damasevicius, and V. H. C. de Albuquerque, ‘‘A novel transfer learning time and computational complexity employing modified CCT for lung
based approach for pneumonia detection in chest X-ray images,’’ Appl. disease classification,’’ Intell. Syst. Appl., vol. 16, 2022, Art. no. 200147.
Sci., vol. 10, no. 2, p. 559, Jan. 2020. [23] A. Mabrouk, R. P. D. Redondo, A. Dahou, M. A. Elaziz, and M. Kayed,
[2] A. U. Ibrahim, M. Ozsoz, S. Serte, F. Al-Turjman, and P. S. Yakoi, ‘‘Pneumonia detection on chest X-ray images using ensemble of deep
‘‘Pneumonia classification using deep learning from chest X-ray images convolutional neural networks,’’ Appl. Sci., vol. 12, no. 13, p. 6448,
during COVID-19,’’ Cognit. Comput., pp. 1–13, Jan. 2021. Jun. 2022.

VOLUME 12, 2024 34705

 M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

[24] Q. Guan, Y. Huang, Z. Zhong, Z. Zheng, L. Zheng, and Y. Yang, ‘‘Diagnose [44] A. Khan, M. U. Akram, and S. Nazir, ‘‘Automated grading of chest X-ray
like a radiologist: Attention guided convolutional neural network for thorax images for viral pneumonia with convolutional neural networks ensemble
disease classification,’’ 2018, arXiv:1801.09927. and region of interest localization,’’ PLoS ONE, vol. 18, no. 1, Jan. 2023,
[25] N. Sri Kavya, T. Shilpa, N. Veeranjaneyulu, and D. Divya Priya, Art. no. e0280352.
‘‘Detecting COVID19 and pneumonia from chest X-ray images using [45] T. Zhou, S. Canu, and S. Ruan, ‘‘Automatic COVID-19 CT segmentation
deep convolutional neural networks,’’ Mater. Today, Proc., vol. 64, using U-Net integrated spatial and channel attention mechanism,’’ Int. J.
pp. 737–743, Jan. 2022. Imag. Syst. Technol., vol. 31, no. 1, pp. 16–27, Mar. 2021.
[26] D. Thakker, V. Shah, J. Rele, V. Shah, and J. Khanapuri, ‘‘Diagnosing child [46] J. Bridge, Y. Meng, Y. Zhao, Y. Du, M. Zhao, R. Sun, and Y. Zheng,
pneumonia using transfer learning,’’ Int. J. Interdiscip. Res. Innov., vol. 7, ‘‘Introducing the GEV activation function for highly unbalanced data
no. 2, pp. 100–104, 2019. to develop COVID-19 diagnostic models,’’ IEEE J. Biomed. Health
[27] S. A. Aljawarneh and R. Al-Quraan, ‘‘Pneumonia detection using Informat., vol. 24, no. 10, pp. 2776–2786, Oct. 2020.
enhanced convolutional neural network model on chest X-ray images,’’ Big [47] D. Chen, S. Ji, F. Liu, Z. Li, and X. Zhou, ‘‘A review of automated diagnosis
Data, Apr. 2023. of COVID-19 based on scanning images,’’ in Proc. 6th Int. Conf. Robot.
[28] M. Shaikh, I. F. Siddiqui, Q. Arain, J. Koo, M. A. Unar, and Artif. Intell., Nov. 2020, pp. 97–104.
N. M. F. Qureshi, ‘‘MDEV model: A novel ensemble-based transfer [48] E. Acar, E. Sahin, and I. Yilmaz, ‘‘Improving effectiveness of different
learning approach for pneumonia classification using CXR images,’’ deep learning-based models for detecting COVID-19 from computed
Comput. Syst. Sci. Eng., vol. 46, no. 1, pp. 287–302, 2023. tomography (CT) images,’’ Neural Comput. Appl., vol. 33, no. 24,
[29] J. A. Prakash, C. Asswin, V. Ravi, V. Sowmya, and K. Soman, ‘‘Pediatric pp. 17589–17609, Dec. 2021.
pneumonia diagnosis using stacked ensemble learning on multi-model [49] D. S. Kermany, M. Goldbaum, W. Cai, C. C. Valentim, H. Liang,
deep CNN architectures,’’ Multimedia Tools Appl., vol. 82, no. 14, S. L. Baxter, A. McKeown, G. Yang, X. Wu, and F. Yan, ‘‘Identifying
pp. 21311–21351, Jun. 2023. medical diagnoses and treatable diseases by image-based deep learning,’’
[30] A. H. Rangkuti, R. Y. Mogot, and V. J. Kusuma, ‘‘A recognizing technique Cell, vol. 172, no. 5, pp. 1122–1131, Feb. 2018.
specific disease on a chest X-ray with support for image clarity and deep [50] H. Bhatt and M. Shah, ‘‘A convolutional neural network ensemble model
learning,’’ Int. J. Intell. Syst. Appl. Eng., vol. 11, no. 3, pp. 176–183, 2023. for pneumonia detection using chest X-ray images,’’ Healthcare Analytics,
vol. 3, Nov. 2023, Art. no. 100176.
[31] Y. Ge, Q. Wang, L. Wang, H. Wu, C. Peng, J. Wang, Y. Xu, G. Xiong,
Y. Zhang, and Y. Yi, ‘‘Predicting post-stroke pneumonia using deep [51] S. Sharma and K. Guleria, ‘‘A systematic literature review on deep
neural network approaches,’’ Int. J. Med. Informat., vol. 132, Dec. 2019, learning approaches for pneumonia detection using chest X-ray images,’’
Art. no. 103986. Multimedia Tools Appl., vol. 83, no. 8, pp. 24101–24151, Aug. 2023.
[52] S. A. Khoiriyah, A. Basofi, and A. Fariza, ‘‘Convolutional neural network
[32] M. Togaçar, B. Ergen, Z. Cömert, and F. Özyurt, ‘‘A deep feature
for automatic pneumonia detection in chest radiography,’’ in Proc. Int.
learning model for pneumonia detection applying a combination of mRMR
Electron. Symp. (IES), Sep. 2020, pp. 476–480.
feature selection and machine learning models,’’ IRBM, vol. 41, no. 4,
pp. 212–222, Aug. 2020. [53] E. Erdem and T. Aydin, ‘‘Detection of pneumonia with a novel CNN-based
approach,’’ Sakarya Univ. J. Comput. Inf. Sci., vol. 4, no. 1, pp. 26–34,
[33] S. Nimbolkar, A. Thakare, S. Mitra, O. Biranje, and A. Sutar, ‘‘Detection
Apr. 2021.
of pneumonia using machine learning and deep learning techniques: An
analytical study,’’ Object Detection by Stereo Vis. Images, vol. 2022, [54] E. Ayan and H. M. Ünver, ‘‘Diagnosis of pneumonia from chest X-ray
pp. 33–55, Sep. 2022. images using deep learning,’’ in Proc. Scientific Meeting Electrical-
Electronics Biomed. Eng. Comput. Sci. (EBBT), Apr. 2019, pp. 1–5.
[34] P. Moeskops, M. A. Viergever, A. M. Mendrik, L. S. de Vries,
[55] K. El Asnaoui, Y. Chawki, and A. Idri, ‘‘Automated methods for detection
M. J. N. L. Benders, and I. Isgum, ‘‘Automatic segmentation of MR brain
and classification pneumonia based on X-ray images using deep learning,’’
images with a convolutional neural network,’’ IEEE Trans. Med. Imag.,
in Artificial Intelligence and Blockchain for Future Cybersecurity Appli-
vol. 35, no. 5, pp. 1252–1261, May 2016.
cations. Berlin, Germany: Springer, 2021, pp. 257–284.
[35] L. Deng and D. Yu, ‘‘Deep learning: Methods and applications,’’ Found.
[56] R. Chiwariro and J. B. Wosowei, ‘‘Comparative analysis of deep
Trends Signal Process., vol. 7, nos. 3–4, pp. 197–387, 2014.
learning convolutional neural networks based on transfer learning for
[36] M. Anthimopoulos, S. Christodoulidis, L. Ebner, A. Christe, and pneumonia detection,’’ Int. J. Res. Appl. Sci. Eng. Technol., vol. 11, no. 1,
S. Mougiakakou, ‘‘Lung pattern classification for interstitial lung diseases pp. 1161–1170, Jan. 2023.
using a deep convolutional neural network,’’ IEEE Trans. Med. Imag., [57] G. Liang and L. Zheng, ‘‘A transfer learning method with deep residual
vol. 35, no. 5, pp. 1207–1216, May 2016. network for pediatric pneumonia diagnosis,’’ Comput. Methods Programs
[37] Z. Yan, Y. Zhan, Z. Peng, S. Liao, Y. Shinagawa, S. Zhang, D. N. Metaxas, Biomed., vol. 187, Apr. 2020, Art. no. 104964.
and X. S. Zhou, ‘‘Multi-instance deep learning: Discover discriminative [58] N. Darapaneni, A. Ranjan, D. Bright, D. Trivedi, K. Kumar, V. Kumar,
local anatomies for bodypart recognition,’’ IEEE Trans. Med. Imag., and A. R. Paduri, ‘‘Pneumonia detection in chest X-rays using neural
vol. 35, no. 5, pp. 1332–1343, May 2016. networks,’’ 2022, arXiv:2204.03618.
[38] S. V. Militante, N. V. Dionisio, and B. G. Sibbaluca, ‘‘Pneumonia detection [59] M. M. Ahsan, K. D. Gupta, M. M. Islam, S. Sen, M. L. Rahman, and
through adaptive deep learning models of convolutional neural networks,’’ M. Shakhawat Hossain, ‘‘Study of different deep learning approach with
in Proc. 11th IEEE Control Syst. Graduate Res. Colloq. (ICSGRC), explainable AI for screening patients with COVID-19 symptoms: Using
May 2020, pp. 88–93. CT scan and chest X-ray image dataset,’’ 2020, arXiv:2007.12525.
[39] I. Sirazitdinov, M. Kholiavchenko, T. Mustafaev, Y. Yixuan, R. Kuleev, and [60] X. Deng, H. Shao, L. Shi, and T. Xie, ‘‘An classification-detection
B. Ibragimov, ‘‘Deep neural network ensemble for pneumonia localization approach of COVID-19 based on chest X-ray and CT by using keras pre-
from a large-scale chest X-ray database,’’ Comput. Electr. Eng., vol. 78, trained deep learning models,’’ Comput. Model. Eng. Sci., vol. 125, no. 2,
pp. 388–399, Sep. 2019. pp. 579–596, 2020.
[40] T. Rahman, M. E. H. Chowdhury, A. Khandakar, K. R. Islam, K. F. Islam, [61] S. Sharma and K. Guleria, ‘‘A deep learning based model for the
Z. B. Mahbub, M. A. Kadir, and S. Kashem, ‘‘Transfer learning with deep detection of pneumonia from chest X-ray images using VGG-16 and neural
convolutional neural network (CNN) for pneumonia detection using chest networks,’’ Proc. Comput. Sci., vol. 218, pp. 357–366, Jan. 2023.
X-ray,’’ Appl. Sci., vol. 10, no. 9, p. 3233, May 2020. [62] Y. Li, Z. Zhang, C. Dai, Q. Dong, and S. Badrigilan, ‘‘Accuracy of deep
[41] A. K. Jaiswal, P. Tiwari, S. Kumar, D. Gupta, A. Khanna, and learning for automated detection of pneumonia using chest X-ray images:
J. J. P. C. Rodrigues, ‘‘Identifying pneumonia in chest X-rays: A deep A systematic review and meta-analysis,’’ Comput. Biol. Med., vol. 123,
learning approach,’’ Measurement, vol. 145, pp. 511–518, Oct. 2019. Aug. 2020, Art. no. 103898.
[42] H. Jung, B. Kim, I. Lee, J. Lee, and J. Kang, ‘‘Classification of lung nodules [63] R. Pramanik, S. Sarkar, and R. Sarkar, ‘‘An adaptive and altruistic PSO-
in CT scans using three-dimensional deep convolutional neural networks based deep feature selection method for pneumonia detection from chest
with a checkpoint ensemble method,’’ BMC Med. Imag., vol. 18, no. 1, X-rays,’’ Appl. Soft Comput., vol. 128, Oct. 2022, Art. no. 109464.
pp. 1–10, Dec. 2018. [64] C. Asswin, K. S. Kumar, A. Dora, V. Ravi, V. Sowmya, E. Gopalakrishnan,
[43] X. Li, L. Shen, X. Xie, S. Huang, Z. Xie, X. Hong, and J. Yu, ‘‘Multi- and K. Soman, ‘‘Transfer learning approach for pediatric pneumonia
resolution convolutional networks for chest X-ray radiograph based lung diagnosis using channel attention deep CNN architectures,’’ Eng. Appl.
nodule detection,’’ Artif. Intell. Med., vol. 103, Mar. 2020, Art. no. 101744. Artif. Intell., vol. 123, Aug. 2023, Art. no. 106416.

34706 VOLUME 12, 2024

M. Ali et al.: Pneumonia Detection Using Chest Radiographs With Novel EfficientNetV2L Model

[65] A. H. Alharbi and H. A. H. Mahmoud, ‘‘Pneumonia transfer learning deep MUHAMMAD FAHEEM MUSHTAQ received
learning model from segmented X-rays,’’ in Proc. MDPI, vol. 10, 2022, the B.S. degree in IT and the M.S. degree in
p. 987. CS from The Islamia University of Bahawalpur,
[66] S. Singh, S. S. Rawat, M. Gupta, B. K. Tripathi, F. Alanzi, Punjab, Pakistan, in 2011 and 2013, respectively,
A. Majumdar, P. Khuwuthyakorn, and O. Thinnukool, ‘‘Deep attention and the Ph.D. degree from the Department of Infor-
network for pneumonia detection using chest X-ray images,’’ Comput., mation Security, Faculty of Computer Science and
Mater. Continua, vol. 74, no. 1, pp. 1673–1691, 2023. Information Technology, University Tun Hussein
[67] X. Xue, S. Chinnaperumal, G. M. Abdulsahib, R. R. Manyam, Onn Malaysia (UTHM), Malaysia, in 2018. He has
R. Marappan, S. K. Raju, and O. I. Khalaf, ‘‘Design and analysis of a
made several contributions through research pub-
deep learning ensemble framework model for the detection of COVID-
lications and book chapters toward information
19 and pneumonia using large-scale CT scan and X-ray image datasets,’’
Bioengineering, vol. 10, no. 3, p. 363, Mar. 2023. security. He received Microsoft certifications of Internet Security and
[68] T. Boyadzhiev, S. Tsvetanov, and S. Dimitrova, ‘‘Deep learning image Acceleration (ISA) Server, Microsoft Certified Professional (MCP), and
classification for pneumonia detection,’’ in Proc. 29th Int. Conf. Syst., Microsoft Certified Technology Professional (MCTS), in 2010. He is
Signals Image Process. (IWSSIP), Jun. 2022, pp. 1–3. currently the Head of the Department of Artificial Intelligence, The Islamia
University Bahawalpur. Previously, he was the Head/an Assistant Professor
with the Department of Information Technology, Khwaja Fareed University
of Engineering and Information Technology, Rahim Yar Khan, Pakistan.
He has been a Research Assistant during the Ph.D. degree, from March
2016 to August 2018. His main research interests include information
security, artificial intelligence, cognitive systems, and applications.
MUDASIR ALI received the B.S.C.S. degree from
the Department of Computer Science, Institute of STEFANÍA CARVAJAL ALTAMIRANDA is currently with Universidad
Southern Punjab, Multan, in 2021. He is currently Europea del Atlántico, Santander, Spain. She is also associated with
pursuing the M.S. degree in computer science Universidade Internacional do Cuanza, Kuito, Bié, Angola and Fundación
with the Department of Computer Science, The Universitaria Internacional de Colombia, Bogotá, Colombia.
Islamia University of Bahawalpur, Bahawalpur,
Pakistan. His current research interests include
the Internet of Things (IoT), artificial intelligence, SILVIA APARICIO OBREGON is currently with Universidad Europea del
data mining, cyber security, machine learning, Atlántico, Santander, Spain. She is also affiliated with the Universidad Inter-
deep learning, and image classification. nacional Iberoamericana, Campeche, Mexico, and Universidad Internacional
Iberoamericana Arecibo, Puerto Rico.

ISABEL DE LA TORRE DÍEZ is currently a

Professor with the Department of Signal The-
ory and Communications and Telematic Engi-
neering, University of Valladolid, Spain, where
MOBEEN SHAHROZ received the M.C.S. degree
she is also the Leader of the GTe Research
and the M.S. degree in computer science from
Group (http://sigte.tel.uva.es). Her research inter-
the Department of Computer Science, Khawaja
ests include the design, development, and evalua-
Fareed University of Engineering and Informa-
tion of telemedicine applications, services and sys-
tion Technology (KFUEIT), Rahim Yar Khan,
tems, e-health, m-health, electronic health records
Pakistan, in 2018 and 2020, respectively. He is
(EHRs), EHRs standards, biosensors, cloud and
currently pursuing the Ph.D. degree in com-
fog computing, data mining, the quality of service (QoS), and the quality
puter science with The Islamia University of
of experience (QoE) applied to the health field.
Bahawalpur. His current research interests include
the Internet of Things (IoT), artificial intelligence,
data mining, natural language processing, machine learning, deep learning, IMRAN ASHRAF received the M.S. degree
and image classification. (Hons.) in computer science from Blekinge Insti-
tute of Technology, Karlskrona, Sweden, in 2010,
and the Ph.D. degree in information and commu-
nication engineering from Yeungnam University,
Gyeongsan-si, South Korea, in 2018. He was a
Postdoctoral Fellow with Yeungnam University,
where he is currently an Assistant Professor with
UROOJ AKRAM received the B.S. degree in computer science from The the Information and Communication Engineering
Islamia University of Bahawalpur, Punjab, Pakistan, in 2013, and the M.S. Department. His research interests include posi-
degree in information technology from the Faculty of Computer Science and tioning using next-generation networks, communication in 5G and beyond,
Information Technology, Universiti Tun Hussein Onn Malaysia (UTHM), location-based services in wireless communication, smart sensors (LIDAR)
Malaysia, in 2018. Her research interests include machine learning and for smart cars, and data analytics.
information security.

VOLUME 12, 2024 34707