Domain adaptation based COVID-19 CT lung infections segmentation network

Han Chen, Yifan Jiang, Hanseok KoarXiv 2020 · arXiv

Abstract

Coronavirus disease (COVID-19 pneumonia) has spread rapidly and become a global epidemic, which has had a great impact on public health and the economy. The automatic segmentation of lung infections from computed tomography (CT) has become an effective method for diagnosis. In order to realize the full potential of deep learning models in COVID-19 pneumonia infections segmentation, a great deal of annotated CT data is needed for training. The data is difficult to collect due to the high risk of infection, and it is laborious to annotate. Recent advances in image synthesis make it possible to train deep learning models on realistic synthetic data with computer-generated annotations. However, the domain shift between real data and synthetic data significantly reduces segmentation performance. In order to solve this issue, we propose a novel domain adaptation based COVID-19 CT lung infections segmentation network. In this work, we use limited real data without annotations and a large amount of annotated synthetic data to train the U-Net segmentation network jointly. To overcome the domain mismatch, we introduce conditional GAN for adversarial training. We update the segmentation network with the cross-domain adversarial loss. This makes the embedding distribution learned by segmentation network from real data and synthetic data closer, thus greatly improving the representation ability of the segmentation network. The experiment results demonstrate that our proposed network significantly outperforms the baseline and state-of-the-art methods.

Introduction

Accurate segmentation of COVID-19 infection areas, such as ground-glass opacities and lung consolidations, in CT scans is vital for rapid diagnosis and severity assessment. However, deep learning models are notoriously data-hungry. Obtaining thousands of pixel-level, expert-annotated medical images is a massive bottleneck due to patient privacy concerns and the grueling manual labor required from radiologists.

While high-quality synthetic CT data with perfect "ground truth" labels can be generated, models trained solely on this data often perform poorly on real patients. This is due to domain shift: subtle differences in texture, noise, and intensity between simulated and real-world scans. This research addresses the challenge of effectively leveraging "free" synthetic labels to improve performance on real-world clinical data where no labels exist.

Proposed Framework

The proposed unsupervised domain adaptation (UDA) framework consists of two main components: a segmentation network and a domain adaptation (DA) module.

  1. The Segmentation Network: Utilizes a standard encoder-decoder structure, comprising a Feature Extractor $f$ and a Pixel-wise Classifier $c$, to predict infection masks.
  2. The Domain Adaptation (DA) Module: This module includes a Generator ($g$) and a Discriminator ($d$). Instead of aligning features directly, the generator reconstructs the original images from internal feature maps. The discriminator then attempts to distinguish between four categories: source-real, source-fake, target-real, and target-fake.
Overall network architecture and domain adaptation module.

A key insight of this approach is that reconstructing images back into "image space" for adversarial training is more robust for medical tasks than direct feature alignment, as medical imaging often contains subtle but critical intensity and texture differences. Furthermore, the generator performs best when utilizing high-level feature maps from the deeper layers of the encoder, suggesting that global semantic information is more useful for domain alignment than low-level pixel details.

Adversarial Training Strategy

The core innovation lies in the cross-domain adversarial training. The network is updated using a loss function that encourages the feature extractor to produce domain-invariant representations. By forcing the model to learn features robust enough that the generator can reconstruct a "source-like" image from a "target" feature map and vice versa, the system aligns the two domains in the latent space.

The training process involves multiple steps to update the generator, discriminator, and finally the feature extractor and classifier. Specifically, in the final update step, the feature extractor is optimized to minimize both the segmentation loss on labeled synthetic data and the adversarial loss, which ensures that the features extracted from real data are indistinguishable from those of synthetic data to the discriminator.

Step 3: Feature extractor and classifier update using adversarial and segmentation losses.

Experimental Results

The model was evaluated using a combination of synthetic and real-world datasets:

  • Source (Synthetic): 10,200 synthetic 2D CT images with pixel-wise annotations for ground-glass opacity (GGO), consolidation, lung, and background.
  • Target (Real): The COVID-19 CT segmentation dataset from the Italian Society of Medical and International Radiology, consisting of 9 CT volumes (~12,000 slices). Training labels for this dataset were withheld during the training phase.

The proposed method significantly outperformed other unsupervised and "source-only" baselines. On the real-world test set, the model achieved a Dice score of 86.15% for multi-class infection segmentation. This represents a nearly 10% absolute improvement over a model trained only on synthetic data (76.98% Dice score). Furthermore, the model outperformed existing state-of-the-art unsupervised methods such as AdvEnt (81.12%) and Self-ensembling (80.43%) across nearly all metrics, including Sensitivity and Specificity. Qualitative success was also noted in the model's ability to capture small, subtle consolidation areas that baseline models often miss.

Conclusion

This work demonstrates that an unsupervised domain adaptation framework can effectively bridge the gap between synthetic and real-world medical data. By leveraging adversarial training in the image reconstruction space, the proposed network achieves state-of-the-art performance without requiring manual expert annotations for the target data. Future work may involve extending the model to 3D volumetric context to utilize the full spatial information available in CT scans and testing on more diverse patient populations to further validate clinical generalizability.

Summary updated

Abstract

Coronavirus disease (COVID-19 pneumonia) has spread rapidly and become a global epidemic, which has had a great impact on public health and the economy. The automatic segmentation of lung infections from computed tomography (CT) has become an effective method for diagnosis. In order to realize the full potential of deep learning models in COVID-19 pneumonia infections segmentation, a great deal of annotated CT data is needed for training. The data is difficult to collect due to the high risk of infection, and it is laborious to annotate. Recent advances in image synthesis make it possible to train deep learning models on realistic synthetic data with computer-generated annotations. However, the domain shift between real data and synthetic data significantly reduces segmentation performance. In order to solve this issue, we propose a novel domain adaptation based COVID-19 CT lung infections segmentation network. In this work, we use limited real data without annotations and a large amount of annotated synthetic data to train the U-Net segmentation network jointly. To overcome the domain mismatch, we introduce conditional GAN for adversarial training. We update the segmentation network with the cross-domain adversarial loss. This makes the embedding distribution learned by segmentation network from real data and synthetic data closer, thus greatly improving the representation ability of the segmentation network. The experiment results demonstrate that our proposed network significantly outperforms the baseline and state-of-the-art methods.
BibTeX
@misc{chen2020domain,
  author = {Han Chen and Yifan Jiang and Hanseok Ko},
  title = {{Domain adaptation based COVID-19 CT lung infections segmentation network}},
  year = {2020},
  eprint = {2011.11242},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2011.11242},
}
Figure