Efficacy of Transfer Learning-based ResNet models in Chest X-ray image classification for detecting COVID-19...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Chest X-ray (CXR) and CT scans are crucial for early COVID-19 detection, especially with non-specific symptoms or negative PCR tests.
• The emergence of new COVID-19 variants increases the diagnostic burden, necessitating advanced methods beyond laboratory testing.
• Deep Learning (DL), particularly Convolutional Neural Networks (CNNs), offers potential for CXR image classification, but faces challenges due to scarce labeled COVID-19 data.

Purpose of the Study:

• To evaluate the performance of various ResNet models in classifying Pneumonia from CXR images.
• To develop and assess a customized ResNet model for improved performance in Pneumonia detection.
• To enhance the reliability of CXR image analysis for clinical decision-making in COVID-19 diagnosis.

Main Methods:

• Utilized Transfer Learning (TL) with pre-trained ResNet architectures (ResNet18_v1 to ResNet152_v1) for CXR image classification.
• Trained and evaluated five standard ResNet models and one customized ResNet model.
• Performed simulations using PyTorch on a Quadro 4000 GPU with detailed system specifications.

Main Results:

• Standard ResNet models achieved global accuracies ranging from 90.87% to 92.95%.
• ResNet50_v1 showed the highest sensitivity (97.18%), ResNet101_v1 the highest specificity (94.02%), and ResNet18_v1 the highest precision (93.53%).
• The customized ResNet model achieved superior performance with 95% global accuracy, 95.65% precision, 92.74% specificity, and 95.9% sensitivity.

Conclusions:

• ResNet models demonstrate significant effectiveness in the automatic detection of Pneumonia from CXR images, supporting COVID-19 diagnosis.
• The developed customized ResNet model offers a reliable tool for analyzing CXR images, aiding clinical decision-making.
• The study highlights the potential of DL and TL approaches to overcome data scarcity issues in medical image analysis for infectious diseases.