Utilizing Transfer Learning and Class-Selective Image Processing for COVID-19 Detection | IEEE Conference Publication | IEEE Xplore

Utilizing Transfer Learning and Class-Selective Image Processing for COVID-19 Detection


Abstract:

The year 2019 marked the emergence of a novel contagious illness named COVID-19 which was brought about by the SARS-CoV-2 virus. This newly identified virus swiftly sprea...Show More

Abstract:

The year 2019 marked the emergence of a novel contagious illness named COVID-19 which was brought about by the SARS-CoV-2 virus. This newly identified virus swiftly spread across the globe, ultimately resulting in the unprecedented COVID-19 pandemic, that impacted populations, healthcare systems, and economies worldwide. Globally, there is a growing need for COVID-19 non-invasive detection by medical image analysis. In the realm of medical diagnostics, deep learning stands out as a supremely efficient approach for analyzing extensive collections of chest X-ray (CXR) images, particularly influencing the realm of COVID-19 screening. A specific technique within deep learning, known as Convolutional Neural Network (CNN), holds a preeminent position due to its effectiveness. In the scope of this research, Convolutional Neural Networks (CNN) have been utilized to differentiate COVID-19 cases from both normal cases and instances of Pneumonia, leveraging information extracted from chest X-ray images. This utilization underscores the potency of advanced machine learning in confronting pivotal healthcare hurdles. The approach involved fine-tuning a pre-trained CNN model, InceptionV3, via transfer learning, wherein certain layers were adjusted before training on the dataset. The initial classification accuracy exhibited strong performance across all image classes, except for one category that displayed a marginally lower accuracy rate. A straightforward yet efficient approach to enhance image contrast has been utilized exclusively in the lower-performance class. This was done to enhance the image quality within that particular class, leading to improved accuracy. Since the entire dataset did not have to go through image processing, the model was computationally efficient and yielded an overall accuracy of 99.92% outperforming existing works.
Date of Conference: 13-15 December 2023
Date Added to IEEE Xplore: 27 February 2024
ISBN Information:
Conference Location: Cox's Bazar, Bangladesh

References

References is not available for this document.