Hybrid classification structures for automatic COVID-19 detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• X-ray images present challenges for disease detection due to low quality and resolution.
• Sophisticated algorithms are necessary for accurate interpretation of medical images.
• Machine learning (ML) and deep learning approaches show promise for improving diagnostic accuracy.

Purpose of the Study:

• To investigate the efficacy of ML and deep learning techniques for COVID-19 detection using X-ray images.
• To compare the performance of manually extracted features versus deep features.
• To evaluate the impact of transfer learning (TL) on COVID-19 detection accuracy.

Main Methods:

• Manual feature extraction followed by comparison of twelve ML classifiers, including Gaussian process (GP) and random forest (RF).
• Deep feature extraction using pre-trained models: ResNet50, ResNet101, Inception-v3, and InceptionResnet-v2.
• Implementation of transfer learning (TL) and comparison with a custom-built convolutional neural network (CNN).

Main Results:

• Gaussian process (GP) and random forest (RF) classifiers showed superiority with manually extracted features.
• InceptionResnet-v2 and ResNet101 models, combined with the GP classifier, yielded optimal performance.
• Deep features and transfer learning (TL) methods achieved the highest accuracy, reaching 100%.

Conclusions:

• Deep features and transfer learning (TL) significantly enhance COVID-19 detection from X-ray images.
• Advanced deep learning models like InceptionResnet-v2 and ResNet101 are highly effective for this task.
• The proposed methods offer a robust and highly accurate solution for automated COVID-19 diagnosis from X-rays.