Contrasting EfficientNet, ViT, and gMLP for COVID-19 Detection in Ultrasound Imagery

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Timely diagnosis of coronavirus is crucial for controlling viral spread.
• Ultrasound imaging offers a potential modality for rapid patient assessment.
• Existing diagnostic methods require timely improvements for effective pandemic management.

Purpose of the Study:

• To develop and evaluate a deep learning-based approach for detecting coronavirus (COVID-19) using ultrasound imagery.
• To compare the performance of EfficientNet-B2, Vision Transformer (ViT), and gMLP models for multiclass classification of ultrasound images.
• To assess the impact of transfer learning on model performance for COVID-19 detection.

Main Methods:

• Utilized transfer learning with an EfficientNet-B2 model pre-trained on ImageNet for ultrasound image classification.
• Compared EfficientNet-B2 against ViT and gMLP models, both with and without transfer learning.
• Employed a multiclass classification strategy to categorize images as COVID-19, pneumonia, or normal.
• Evaluated models on a public ultrasound dataset comprising 261 recordings from 216 patients.

Main Results:

• EfficientNet-B2 with transfer learning achieved the highest performance, with precision, recall, and F1 scores of 95.84%, 99.88%, and 97.41% for the COVID-19 class, respectively.
• The best model demonstrated an overall accuracy of 96.79%.
• EfficientNet-B2 with transfer learning outperformed ViT (92.82% accuracy) and gMLP (93.03% accuracy).

Conclusions:

• Transfer learning, particularly with EfficientNet-B2, significantly enhances the accuracy of deep learning models for COVID-19 detection in ultrasound images.
• Deep learning models show promise for rapid and accurate COVID-19 diagnosis, aiding in infection control.
• EfficientNet-B2 is a highly effective architecture for classifying ultrasound images for respiratory diseases including COVID-19.