[Paper Review] TUNA-Net: Task-oriented UNsupervised Adversarial Network for Disease Recognition in Cross-Domain Chest X-rays

Yuxing Tang, Youbao Tang+3 more|arXiv (Cornell University)|Aug 21, 2019

COVID-19 diagnosis using AI18 references4 citations

TL;DR

TUNA-Net is a task-oriented, unsupervised adversarial network that enables cross-domain pneumonia recognition in chest X-rays by preserving class-specific semantics during cycle-consistent image-to-image translation. It achieves 96.3% AUC on pediatric X-rays—nearly matching the 98.1% AUC of a fully supervised model—without requiring target-domain labels.

ABSTRACT

In this work, we exploit the unsupervised domain adaptation problem for radiology image interpretation across domains. Specifically, we study how to adapt the disease recognition model from a labeled source domain to an unlabeled target domain, so as to reduce the effort of labeling each new dataset. To address the shortcoming of cross-domain, unpaired image-to-image translation methods which typically ignore class-specific semantics, we propose a task-driven, discriminatively trained, cycle-consistent generative adversarial network, termed TUNA-Net. It is able to preserve 1) low-level details, 2) high-level semantic information and 3) mid-level feature representation during the image-to-image translation process, to favor the target disease recognition task. The TUNA-Net framework is general and can be readily adapted to other learning tasks. We evaluate the proposed framework on two public chest X-ray datasets for pneumonia recognition. The TUNA-Net model can adapt labeled adult chest X-rays in the source domain such that they appear as if they were drawn from pediatric X-rays in the unlabeled target domain, while preserving the disease semantics. Extensive experiments show the superiority of the proposed method as compared to state-of-the-art unsupervised domain adaptation approaches. Notably, TUNA-Net achieves an AUC of 96.3% for pediatric pneumonia classification, which is very close to that of the supervised approach (98.1%), but without the need for labels on the target domain.

Motivation & Objective

• Address the challenge of domain shift in medical image classification when transferring models from labeled source domains (e.g., adult X-rays) to unlabeled target domains (e.g., pediatric X-rays).
• Overcome limitations of existing unsupervised domain adaptation methods that fail to preserve class-specific lesion semantics during image translation.
• Develop a framework that maintains low-level image details, mid-level features, and high-level semantic content during cross-domain image translation to support downstream disease recognition.
• Enable effective disease recognition in target domains without requiring manual annotation of the target data, reducing clinical labeling burden.

Proposed method

• Propose a cycle-consistent generative adversarial network (Cycle-GAN variant) with task-driven, discriminatively trained components to align source and target domains.
• Integrate class-aware semantic constraint modules that use source domain labels to guide image translation while preserving disease-specific content.
• Train a joint optimization framework that simultaneously minimizes domain discrepancy, reconstructs real images from translated ones, and regularizes mid-level features across domains.
• Use a dual-pathway classification head: one for source domain (with labels) and one for target domain (trained end-to-end with synthetic images), enabling online adaptation.
• Incorporate a feature construction loss to align mid-level features between real and synthetic target images, improving generalization.
• Apply adversarial training to ensure synthesized target-domain images are both anatomically realistic and semantically faithful to the original disease patterns.

Experimental results

Research questions

• RQ1 Can unsupervised domain adaptation methods preserve class-specific disease semantics during cross-domain image translation in medical imaging?
• RQ2 How does incorporating source-domain labels into the image translation process affect downstream disease classification performance on an unlabeled target domain?
• RQ3 To what extent can a cycle-consistent GAN framework with task-specific constraints outperform state-of-the-art UDA methods in cross-domain chest X-ray pneumonia recognition?
• RQ4 Does end-to-end joint training of the target classifier with the translation network improve performance compared to offline training or ablated components?

Key findings

• TUNA-Net achieves an AUC of 96.3% on pediatric chest X-ray pneumonia classification, which is very close to the 98.1% AUC of a fully supervised model trained on target-domain labels.
• The model outperforms state-of-the-art unsupervised domain adaptation methods such as ADDA (91.8% AUC) and CyCADA (93.5% AUC) under identical settings.
• Ablation studies show that online end-to-end training of the target classifier is the most critical component, contributing most to performance improvement.
• The inclusion of feature construction loss and reconstruction of source images from synthetic target images significantly enhances model robustness and generalization.
• Qualitative results demonstrate that TUNA-Net preserves pathological features (e.g., lobar consolidation) during translation, while Cycle-GAN and CyCADA often obscure or lose such details.
• Synthesized images from TUNA-Net are both anatomically realistic and semantically faithful to the source input, especially in preserving disease content.