[Paper Review] COVID-DA: Deep Domain Adaptation from Typical Pneumonia to COVID-19

Yifan Zhang, Shuaicheng Niu+7 more|arXiv (Cornell University)|Apr 30, 2020

COVID-19 diagnosis using AIMedicine52 references57 citations

TL;DR

COVID-DA introduces a domain-adaptive framework that transfers knowledge from typical pneumonia to COVID-19 diagnosis by disentangling domain-shared and domain-specific classifiers and aligning feature and joint distributions.

ABSTRACT

The outbreak of novel coronavirus disease 2019 (COVID-19) has already infected millions of people and is still rapidly spreading all over the globe. Most COVID-19 patients suffer from lung infection, so one important diagnostic method is to screen chest radiography images, e.g., X-Ray or CT images. However, such examinations are time-consuming and labor-intensive, leading to limited diagnostic efficiency. To solve this issue, AI-based technologies, such as deep learning, have been used recently as effective computer-aided means to improve diagnostic efficiency. However, one practical and critical difficulty is the limited availability of annotated COVID-19 data, due to the prohibitive annotation costs and urgent work of doctors to fight against the pandemic. This makes the learning of deep diagnosis models very challenging. To address this, motivated by that typical pneumonia has similar characteristics with COVID-19 and many pneumonia datasets are publicly available, we propose to conduct domain knowledge adaptation from typical pneumonia to COVID-19. There are two main challenges: 1) the discrepancy of data distributions between domains; 2) the task difference between the diagnosis of typical pneumonia and COVID-19. To address them, we propose a new deep domain adaptation method for COVID-19 diagnosis, namely COVID-DA. Specifically, we alleviate the domain discrepancy via feature adversarial adaptation and handle the task difference issue via a novel classifier separation scheme. In this way, COVID-DA is able to diagnose COVID-19 effectively with only a small number of COVID-19 annotations. Extensive experiments verify the effectiveness of COVID-DA and its great potential for real-world applications.

Motivation & Objective

• Motivate diagnosing COVID-19 from chest radiographs when COVID-19 annotations are scarce.
• Address both domain distribution mismatch and task difference between pneumonia and COVID-19.
• Develop a deep learning model that leverages labeled pneumonia data to improve COVID-19 diagnosis with limited target labels.

Proposed method

• Use a domain-shared feature extractor to learn domain-invariant representations.
• Incorporate two domain discriminators to align feature distributions and joint feature-prediction distributions.
• Introduce a classifier separation scheme with a domain-shared classifier and two domain-specific classifiers (one per domain).
• Maximize diversity between domain-shared and domain-specific classifiers to capture domain-private information.
• Train with a focal loss to handle label imbalance and improve discrimination.

Experimental results

Research questions

• RQ1 Can domain adversarial learning align features across pneumonia and COVID-19 images despite domain shifts?
• RQ2 Does separating classifiers into shared and domain-specific components improve COVID-19 diagnosis with limited target labels?
• RQ3 What is the impact of explicitly maximizing diversity between classifiers on cross-domain diagnosis performance?
• RQ4 How does COVID-DA perform compared with baselines and other domain adaptation methods on COVID-19 diagnosis?
• RQ5 Is the method robust to imbalanced classes common in medical imaging datasets?

Key findings

• COVID-DA achieves the highest F1 of 92.98% and AUC of 0.985 on COVID-19 diagnosis among evaluated methods.
• COVID-DA outperforms Source-only, Target-only, Fine-tuning, standard DA methods (MCD, DANN, DSN, DMAN), and semi-supervised DA methods (SDT, semi-DMAN).
• Ablation studies show all components (feature adversarial adaptation, classifier adversarial adaptation, classifier diversity, and focal loss) contribute to performance gains.
• Visual Grad-CAM analysis indicates domain-shared and target-specific classifiers focus on complementary regions for diagnosis.
• The ensemble of domain-shared and domain-specific classifiers improves robustness and interpretability for clinical use.