Multiple-Inputs Convolutional Neural Network for COVID-19 Classification and Critical Region Screening From Chest...

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Radiology

Background:

• The COVID-19 pandemic presents a significant global health challenge.
• Chest X-ray radiography (CXR) is crucial for diagnosing COVID-19.
• Manual feature extraction from CXRs is labor-intensive for radiologists.

Purpose of the Study:

• To develop a novel multiple-inputs convolutional neural network (MI-CNN).
• To classify COVID-19 and extract critical regions from CXRs.
• To evaluate the impact of input numbers on MI-CNN performance.

Main Methods:

• Utilized 6205 CXR images (3021 COVID-19, 3184 normal).
• Segmented CXRs into 2, 4, or 16 regions, each serving as an MI-CNN input.
• Fused CNN features from multiple inputs for COVID-19 classification.
• Assessed region contributions to classification accuracy.

Main Results:

• MI-CNNs showed high efficiency in classifying COVID-19 CXRs.
• Increased inputs (2, 4, 16) improved MI-CNN performance over single-input CNNs.
• Left- and right-lung ROIs showed ~4% lower accuracy than whole-image analysis.
• Right-lung regions were more contributory to COVID-19 classification; left-lung regions to normal classification.

Conclusions:

• MI-CNN accuracy increases with more inputs, particularly the 16-input model.
• This approach aids radiologists in COVID-19 CXR identification.
• The method facilitates screening of critical regions associated with COVID-19.