RCoNet: Deformable Mutual Information Maximization and High-Order Uncertainty-Aware Learning for Robust COVID-19...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Radiology and Diagnostic Imaging
• Infectious Disease Diagnostics

Background:

• COVID-19 poses a global healthcare challenge, with chest X-rays (CXRs) being a preferred diagnostic tool due to speed and cost.
• Challenges in CXR-based COVID-19 detection include distinguishing it from pneumonia and handling noisy data, leading to misclassifications.
• Existing methods often prioritize prediction accuracy over crucial uncertainty estimation, especially with limited or noisy training data.

Purpose of the Study:

• To develop a robust deep learning network for accurate and reliable COVID-19 detection from chest X-ray images.
• To address the limitations of existing methods by incorporating uncertainty estimation and improving feature representation.
• To enhance the performance of COVID-19 detection in the presence of data noise and ambiguity.

Main Methods:

• Proposed RCoNet^k_s, a novel deep network incorporating Deformable Mutual Information Maximization (DeIM), Mixed High-order Moment Feature (MHMF), and Multiexpert Uncertainty-aware Learning (MUL).
• DeIM was used to maximize mutual information for compact and disentangled feature representations.
• MHMF explored high-order statistics for discriminative feature extraction, while MUL employed parallel dropout networks for uncertainty evaluation.

Main Results:

• RCoNet^k_s achieved state-of-the-art performance on a large COVIDx dataset (15,134 CXR images).
• The method demonstrated superior effectiveness compared to existing techniques, particularly in the presence of noisy data.
• Experimental results confirmed improved accuracy and robustness in COVID-19 detection using CXR images.

Conclusions:

• RCoNet^k_s offers a robust and accurate solution for COVID-19 detection using chest X-rays.
• The integration of DeIM, MHMF, and MUL effectively addresses challenges related to feature ambiguity and data noise.
• This approach enhances diagnostic reliability and provides valuable uncertainty estimation for clinical decision-making.