Page 38 - AIH-2-1

Artificial Intelligence in Health Deep learning on chest X-ray and CT for COVID-19

(iii) peribronchial rounded consolidations, (iv) multifocal learning models in improving diagnostic accuracy and
bilateral consolidations, (v) ball pattern or round pneumonia, early disease detection.
and (vi) bilateral symmetrical diffuse lung involvement. Out In this study, we developed deep transfer learning-based
of these patterns, pattern (vi) is the most severe, suggesting approaches for the automatic detection of COVID-19 using
acute respiratory distress syndrome in the patient. various CNN-based models on chest X-ray images and

Several research groups have used deep learning‐based implemented several innovations in the training protocol.
techniques for COVID‐19 and pneumonia detection. 15-31 Beyond the immediate applicability of the present method
Since most of these techniques are well discussed and as an additional diagnostic tool for identifying COVID-19
debated in the literature, we will very briefly review only infections, it also might have far-reaching consequences. An
a few of them. Wang et al. used deep learning techniques earlier big challenge was to identify and detect the infected
16
on CT images to screen COVID‐19 patients with an individual; however, slowly but steadily, the major focus
accuracy, specificity, and sensitivity of 89.5%, 88%, and has shifted toward the long-term care of the lungs of the
87%, respectively. A novel convolutional neural network infected persons. All seven human coronaviruses affect the
(CNN), COVID‐Net for detecting COVID‐19 using chest respiratory tracts and lungs. However, three of these virus
X‐ray images presented by a different research team has types, i.e., SARS-CoV, Middle East respiratory syndrome
an accuracy of 83.5%. Transfer learning offers several coronavirus, and SARS-CoV-2 are known to severely affect
15
benefits, primarily in saving training time, significantly the lungs. Unfortunately, the long-term impact of SARS-
improving the performance of ML models, and requiring CoV-2 infection is neither clearly understood nor widely
smaller data sets for training. For example, Sohaib et al. studied as the viruses are dynamically evolving across the
8
proposed a novel approach (demonstrated in non- world. Most severe infections are known to cause post-
medical applications) that used step transfer learning COVID-19 sequelae, i.e., fibrosis of the lungs in the future.
(STL) combined with extreme learning machine (ELM), A worthwhile initiative toward recording post-COVID-19
primarily focusing on improving the generalization power symptoms in this direction has been reported in the
of deep learning models for autonomous inspection. By literature. An ML-based system that relies on the chest
33
leveraging STL to extract generalized abstract features X-ray and/or CT scan image analysis might prove to be
from diverse source images and utilizing ELM to overcome highly valuable and instrumental in the long-term care of
optimization limitations of traditional neural networks, the lungs by appropriately creating a highly scalable and
their model achieved significant improvements in accuracy easily integrable digital infrastructure through curating,
(2.5%), recall (4.8%), and precision (0.8%) compared to cataloging, classifying and most importantly, creating an
existing studies. This approach enhanced generalization appropriate data repository of the vital information about
demonstrating the usefulness of transfer learning different stages of the lungs as a function of time (using
techniques for increasing the robustness of the detection automated time-stamps), significantly augmenting the
models and making the model more generalizable. Transfer healthy lung initiative in the post-COVID-19 era.
learning has found many applications in the biomedical
arena. Joaquin used a small dataset of 339 images for 2. Data and methods
17
32
training and testing by utilizing the ResNet50-based deep 2.1. Data
transfer learning technique and obtained a validation
accuracy of 96.2%. Ahmmed et al. conducted an in-depth The dataset of chest X-ray and CT scan images 34,35 used
32
analysis of brain tumor classification using transfer learning in this study were sourced from a freely available GitHub
36
across multiple classes, utilizing robust frameworks, repository maintained by Cohen. X-ray images containing
such as ResNet 50 and Inception V3 for MRI images. pneumonia and normal images were obtained from the
37
Their research meticulously curated paired datasets and Kaggle dataset. We used 1763 images to analyze the
incorporated advanced techniques such as Early Stopping, performance of the neural networks used for this study and
ReduceLROnPlateau, and hyperparameter optimization. utilized 1260 images to train the model along with 251 and
These strategies significantly improved model accuracy, 252 images for validating and testing the model, respectively.
achieving exceptional classification rates for various types The dataset consists of 563, 947, and 223 images that belong to
of brain tumors. Similarly, Podder et al. developed a deep COVID-19, normal, and pneumonia categories, respectively.
31
learning model using optimized DenseNet architectures to Figure 1 shows some examples from the dataset.
diagnose infectious diseases from chest X-rays, achieving
high detection rates for COVID-19 and other conditions. 2.2. Methods: Different CNN architectures
Their modifications to the DenseNet architecture and Various CNN models such as ResNet, SeResNext, DenseNet,
hyperparameter tuning demonstrated the potential of deep and EfficientNet were used for classification. Here we briefly

Volume 2 Issue 1 (2025) 32 doi: 10.36922/aih.2888