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

Yuxing Tang, Youbao Tang, Veit Sandfort, Jing Xiao, Ronald M. Summers

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. 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.

Introduction

Deep learning models in radiology often suffer from "domain shift." A model trained on X-rays from one hospital or age group (e.g., adults) frequently performs poorly when applied to another (e.g., children) due to differences in anatomy, imaging protocols, and hardware. While supervised learning can mitigate this, labeling medical data is expensive and requires significant expert clinician time.

Unsupervised Domain Adaptation (UDA) aims to solve this by translating knowledge from a labeled source domain to an unlabeled target domain. However, existing UDA methods, such as standard Cycle-GANs, often focus on matching the general "style" of images while inadvertently erasing critical disease features (like lung opacities). TUNA-Net addresses this by integrating task-specific semantic constraints into the translation process.

Methodology

TUNA-Net (Task-oriented UNsupervised Adversarial Network) builds upon the Cycle-GAN architecture, which learns to translate images between two domains (Source $A$ $\leftrightarrow$ Target $P$) without paired examples. To ensure disease information is preserved, the framework incorporates three layers of constraints:

  1. Low-level Detail Preservation: Utilizes cycle-consistency loss to ensure translated images can be reconstructed back to their original state.
  2. High-level Semantic Modeling: Auxiliary classifiers ($F_A$ and $F_P$) force generators to maintain disease features. For instance, if an adult X-ray has pneumonia, its synthetic pediatric counterpart must also display pneumonia features.
  3. Mid-level Feature Regularization: This ensures that the internal feature representations for real and synthetic target images are statistically similar, preserving structural representations common to both domains.
The TUNA-Net architecture integrates task-specific constraints into a cycle-consistent GAN framework.

Experimental Setup

The study utilized two major public chest X-ray datasets:

  • Source Domain (Adults): A subset of the NIH Chest X-ray 14 dataset, containing 6,993 normal and 4,659 pneumonia images.
  • Target Domain (Pediatric): Data from the Guangzhou Women and Children’s Medical Center. 5,232 images were used for training (with labels hidden) and 624 images for evaluation.

Results and Analysis

TUNA-Net significantly outperformed standard UDA baselines. While a standard Cycle-GAN achieved only an 80.4% AUC—performing worse than no adaptation (89.3%) due to "semantic washing"—TUNA-Net reached a 96.3% AUC. This performance is remarkably close to the supervised upper bound of 98.1%.

Quantitative comparison of TUNA-Net against baselines and ablation studies.

Ablation studies (shown in the results table) confirm that both the feature loss and the auxiliary classifiers are vital. Furthermore, training the target classifier ($F_P$) "online" during the translation process was found to be more effective than "offline" training on pre-generated synthetic images.

Conclusion

TUNA-Net successfully bridges the domain gap between adult and pediatric radiology datasets without requiring target-domain labels. By enforcing task-oriented constraints, it prevents the loss of clinical semantics common in general-purpose GANs. Future work will explore extending this framework to multi-label classification and more diverse imaging modalities like CT and MRI.

Summary updated

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.
BibTeX
@inproceedings{tang2019tuna,
  author = {Yuxing Tang and Youbao Tang and Veit Sandfort and Jing Xiao and Ronald M. Summers},
  title = {{TUNA-Net: Task-oriented UNsupervised Adversarial Network for Disease Recognition in Cross-Domain Chest X-rays}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2019},
  url = {https://openreview.net/forum?id=F5IpWeDniC},
}
Figure