Uncertainty-aware convolutional neural network for COVID-19 X-ray images classification

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Radiology

Background:

• Deep learning (DL) models have advanced medical image analysis, including COVID-19 detection from chest X-rays (CXRs).
• Existing DL models lack inherent uncertainty estimation, a critical feature for medical diagnostics.

Purpose of the Study:

• To develop an uncertainty-aware convolutional neural network (UA-ConvNet) for automated COVID-19 detection from CXRs.
• To estimate the uncertainty associated with DL model predictions in medical imaging.

Main Methods:

• Fine-tuning the EfficientNet-B3 model on CXR datasets.
• Employing Monte Carlo (MC) dropout during inference for multiple forward passes.
• Calculating mean and entropy from the predictive distribution to quantify uncertainty.

Main Results:

• The UA-ConvNet achieved a G-mean of 98.02% and 99.16% for multi-class and binary classification, respectively, on different datasets.
• Sensitivity reached 98.15% (multi-class) and 99.30% (binary).
• The model demonstrated superior performance compared to existing methods in diagnosing COVID-19 from CXRs.

Conclusions:

• The proposed UA-ConvNet effectively detects COVID-19 from CXRs while providing crucial uncertainty estimates.
• This approach enhances the reliability of AI in medical diagnostics.
• The method shows significant potential for improving automated screening of infectious diseases.