Multi-attention representation network partial domain adaptation for COVID-19 diagnosis

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Machine Learning for Healthcare

Background:

• Accurate COVID-19 diagnosis is crucial for controlling its spread, but current methods require extensive labeled data.
• Domain adaptation (DA) offers a solution by leveraging existing pneumonia datasets, but faces challenges like negative transfer and noisy data.
• Existing partial domain adaptation (PDA) methods learn a single feature representation, limiting their ability to capture comprehensive image information.

Purpose of the Study:

• To propose a novel multi-attention representation network partial domain adaptation (MARPDA) model to address the limitations of current DA and PDA methods for COVID-19 diagnosis.
• To enhance the accuracy and reliability of computer-aided diagnosis for COVID-19 by effectively handling noisy data and learning from multiple feature representations.
• To improve upon state-of-the-art methods in classifying COVID-19 and pneumonia using limited labeled target domain data.

Main Methods:

• Developed MARPDA, incorporating multiple attention-based representation networks to learn from diverse feature spaces.
• Implemented a sample-weighted strategy to facilitate partial data transfer and mitigate negative transfer from noisy source data.
• Applied the MARPDA model to classify pneumonia and COVID-19 using chest imaging data.

Main Results:

• MARPDA achieved higher classification accuracy for COVID-19 diagnosis compared to existing state-of-the-art methods.
• The model demonstrated stability and reliability, validated through confusion matrix analysis and performance curve experiments.
• The multi-attention approach effectively learned fine-grained image features from multiple representations, outperforming single-representation methods.

Conclusions:

• MARPDA offers a robust and effective solution for COVID-19 diagnosis, particularly in scenarios with limited labeled data and noisy training sets.
• The proposed method significantly advances the field of domain adaptation for medical image analysis, providing a more accurate and reliable diagnostic tool.
• MARPDA's ability to learn from multiple feature spaces and handle noisy data makes it a promising approach for real-world clinical applications in infectious disease detection.