DCML: Deep contrastive mutual learning for COVID-19 recognition.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• Computer-aided diagnosis of COVID-19 is crucial for efficient and rapid identification.
• Existing deep learning models often focus on network depth or width, neglecting inter-sample relationships.
• The potential of contrastive learning in enhancing COVID-19 recognition remains underexplored.

Purpose of the Study:

• To propose a novel Deep Contrastive Mutual Learning (DCML) model for improved COVID-19 diagnosis.
• To leverage contrastive learning within a deep mutual learning framework to better utilize sample relationships.
• To enhance feature discriminability and model generalizability for COVID-19 detection.

Main Methods:

• Implemented a multi-way data augmentation strategy using Fast AutoAugment (FAA) to expand the training dataset and mitigate overfitting.
• Integrated contrastive learning principles into the Deep Mutual Learning (DML) framework to analyze relationships between diverse samples.
• Developed an adaptive model fusion technique to generate more discriminative image features.

Main Results:

• The proposed DCML model demonstrated superior performance compared to state-of-the-art methods across three public datasets.
• DCML effectively mines implicit contrastive relationships between samples, leading to more robust feature extraction.
• The model achieved higher accuracy in COVID-19 recognition, validating its diagnostic capabilities.

Conclusions:

• DCML offers a practical and efficient approach to automatic COVID-19 recognition.
• The integration of contrastive learning significantly enhances the performance of deep mutual learning for medical image analysis.
• The proposed method shows high potential for clinical application due to its improved accuracy and reproducibility.