Multi-feature Multi-Scale CNN-Derived COVID-19 Classification from Lung Ultrasound Data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• The COVID-19 pandemic strained healthcare systems globally.
• Computed tomography (CT) and chest X-ray (CXR) are primary imaging tools but have limitations.
• Limitations include transmission risk, radiation exposure, and cost-effectiveness.

Purpose of the Study:

• To evaluate lung ultrasound (LUS) as a viable alternative for COVID-19 diagnosis.
• To develop and assess a deep learning model for COVID-19 classification using LUS data.
• To explore image processing and feature fusion techniques for improved diagnostic accuracy.

Main Methods:

• Utilized local phase filtering and radial symmetry transformation on LUS images.
• Employed a multi-scale residual convolutional neural network (CNN) for classification.
• Investigated image combination and multi-level feature fusion strategies.

Main Results:

• The deep learning model demonstrated promising performance in COVID-19 prediction.
• Evaluated on POCUS and ICLUS-DB datasets, showing effective classification capabilities.
• Objective diagnostic information was generated for clinicians.

Conclusions:

• Lung ultrasound, enhanced by deep learning, provides a practical and accurate method for COVID-19 assessment.
• The proposed multi-scale CNN with feature fusion shows potential for objective COVID-19 diagnosis.
• LUS offers advantages over traditional methods like CT and CXR for COVID-19 screening.