Socio-economic Challenges in COVID Detection using Transfer

Learning-Based Methods

DITJONA KULE1, OGERTA ELEZAJ2, UMESH MEHTRE2

ALBANIA

period (WHO, 2016). The core of innovation in the healthcare industry lies in automation systems, particularly in

the realm of image detection. As the ratio of healthcare workers to patients decreases, the integration of robotics

and artificial intelligence plays a crucial role in bridging the gap. These technologies not only compensate for the

declining workforce but also bring a level of accuracy and precision that eliminates the potential for human error

in image detection processes. In this paper we focus on the COVID-19 pandemic that presents significant

socio-economic challenges, impacting various aspects of daily life, including health, the economy, and social

development. The need for chest X-ray (CXR) scans is rising due to pneumonia being a critical and common

complication of COVID-19. Early detection and diagnosis are pivotal in curbing the spread of the virus,

prompting the utilization of the reverse transcription polymerase chain reaction (RT-PCR) as the predominant

screening technology. Nevertheless, the task's complexity, time-consuming nature, and reported insensitivity in

classification performance with an accuracy of 97.33%.

Key-Words: - Deep learning, COVID-19, Chest x-ray, Transfer learning, Image processing, Explainable artificial

intelligence.

Received: June 16, 2023. Revised: February 5, 2024. Accepted: March 13, 2024. Published: May 7, 2024.

Shortages of staff in European radiology departments

other functions. Numerous AI applications in

medical imaging undergo training using extensive

preserved each year, showcasing the profound

impact on public health and well-being.

Financial Savings: The initiative generates

significant economic advantages, with savings

direct monetary savings and accounts for opportunity

effectiveness in various sectors. The impact on

healthcare professionals, particularly on healthcare

professionals' physical and emotional well-being, is

noteworthy.

rapidly. Since 2005, WHO has declared severe

pandemics as PHEIC. Six cases, including H1N1 in

2009 and COVID-19 in 2020, have been declared,

[1]. Over 599 million COVID-19 cases have been

found globally, with almost 6.46 million deaths, [2].

Radiologists can identify these symptoms as signs of

lung issues and respiratory infections, [3]. RT-PCR,

also called real-time polymerase chain reaction, is

the most effective method for detecting COVID-19

cases, [4]. However, RT-PCR kits can be costly and

can take six to nine hours to confirm a diagnosis. The

poor sensitivity of RT-PCR leads to a high

percentage of observations that are incorrectly

interpreted as negative. It has been determined that

radiological imaging techniques, such as chest

can be easily recognized using chest X-rays as a

diagnostic tool. Therefore, radiologists are required

to investigate these signals very carefully. It is a hard

endeavor that requires a significant amount of time

investment.

The clinical tests that are used to detect COVID-19

can be quite pricey and time-consuming, which

Deep learning techniques offer a more

Architectural Evaluation: The study evaluates

VGG16, InceptionResNetV2, and ResNet50,

architectures to analyze and differentiate between

COVID-19 and healthy patients.

Data Handling Techniques: It implements image

follows. Section 2 presents related work on using

deep learning methods for COVID-19 detection from

x-ay images. The methodology used is introduced in

Sect. 3. Section 4 describes the results obtained by all

the experiments, and section 5 sheds light on the

discussion of the results. Finally, the conclusions and

future work are given in Sect. 6.

AI is set to revolutionize the interpretation of medical

images in digital pathology, encompassing the

X-rays are pivotal, and AI algorithms in digital

pathology prove to be valuable by autonomously

detecting pathologies, potentially outperforming

radiologists in accuracy. This advancement has the

potential to save up to 1,900 lives annually.

AI excels in early detection. Utilizing machine

learning on coronary computed tomographic

angiography images and clinical data, algorithms

demonstrate superior accuracy in predicting

breast cancer. Despite breast cancer remaining a

leading cause of death among EU women, with

around 85,000 annual deaths, early detection is key

to successful treatment. AI has shown substantial

benefits in mammography screening, significantly

Throughout the European Union, a shortage of

Authors in [11], introduced a modern

X-ray (CXR) images. They have been using

different methods and the results of all these

methods have been evaluated and compared. Firstly,

during the data pre-processing phase, all the images

in the dataset were transformed to a resolution of

224 by 224 pixels. In the second stage, the image

network classifiers. No augmentation techniques

were applied to the images. The results of the

experiments concluded that the VGG19 model

outperformed the other models, achieving the

highest accuracy rate at 90%.

the Decompose, Transfer, and Compose (DeTraC)

aiming at diagnosing COVID by the images, with

the primary goal of identifying abnormalities within

the images. To substantiate their findings, the

These results underscore the effectiveness of the

DeTraC for the classification of COVID-19 chest

X-ray images, showcasing its potential to identify

abnormalities with a high level of accuracy.

samples, with 2,143 cases of COVID-19, 3,674

instances of pneumonia (excluding COVID-19), and

4,223 healthy cases (neither COVID-19 nor

pneumonia). This model achieved an accuracy of

96.43% and a sensitivity of 93.68%, [13]. They

fine-tuning scheme for the pre-trained ResNet-50

architecture, creating the COVIDResNet model.

This model achieved an accuracy of 96.23%. The

multi-stage fine-tuning approach likely involved

adjusting the weights of the pre-trained ResNet-50

layers to enhance the model's performance on a

specific task, in this case, the classification of

Authors in [15], used a deep convolutional

Table 1. Accuracy of the models using X-ray images

normal and COVID-19 pneumonia, along with

normal, viral, and COVID-19 pneumonia with and

without image enhancement. The dataset also

included normal and viral pneumonia samples. Four

X-ray images. They employed Inception V3 with

transfer learning to identify infections in the

patient's chest. The model underwent testing on a

dataset comprising 1341 normal, 1345 viral

pneumonia, and 864 COVID-19 images. The

classification accuracy of the model reached 96%.

highlighting its efficacy in accurately identifying

COVID-19 cases. This result signifies the potential

applicability of Inception-ResNetV2 in diagnostic

processes for COVID-19 through chest X-ray

images. Table 1 summarizes the key accuracy

metrics discussed.

and 1,345 Viral Pneumonia pictures. However, our

In this section, we present an analysis of the key

findings derived from each investigation conducted

in this study. We provide an overview of the

pre-trained and sequential models utilized, along

Table 3. Comparison of the accuracy for the

sequential model balanced and unbalanced dataset

Fig. 4: Accuracy and Loss of VGG19 using

sequential model

5, offering valuable insights into the evolution of

each model's accuracy and loss throughout the

training process.

This section explores pre-trained models, including

VGG19, InceptionResNetV2, and ResNet-50,

utilizing binary cross-entropy to distinguish between

COVID and normal images. The dataset was split

classifiers utilized activation="sigmoid" with global

average pooling 2D and binary cross-entropy before

balanced dataset, VGG19 achieved the highest score

78.85%. In the unbalanced dataset,

InceptionResNetV2 yielded results comparable to

VGG19, scoring 73, while ResNet-50 scored

50.92%.

optimizer. Over 10 epochs, the learning rate for all

pre-trained models was set at 0.0001 per iteration.

VGG19 exhibited the highest performance, scoring

98.70%, followed by InceptionResNetV2 (97.51%),

and ResNet-50 (96.64%). Similarly, on the

Fig. 9: Accuracy and Loss of InceptionResNetV2

using Focal loss

The performance metrics of using Focal loss for

VGG19 are shown in Figure 8, while those of

InceptionResNetV2 are illustrated in Figure 9,

offering valuable insights into the evolution of each

model's accuracy and loss throughout the training

process.

augmentation strategies and focal loss, to enhance

the performance of transfer learning methods. The

results, presented in Tables 4 and 5, reveal that

InceptionResNetV2 stands out with a remarkable

accuracy of 99.52% when dealing with an

unbalanced dataset. The study introduces three

and InceptionResNetV2—used for classifying

COVID-19 and normal cases from chest X-ray

radiography images. Across balanced datasets,

Notable examples include COVIDX-Net

As shown in Table 5, based on our evenly