Symmetrical awareness network for cross-site ultrasound thyroid nodule segmentation

Wenxuan Ma, Xiaopeng Li, Lian Zou, Cien Fan, Meng WuFrontiers in Public Health 2023 · DOI

Abstract

Recent years have seen remarkable progress of learning-based methods on Ultrasound Thyroid Nodules segmentation. However, with very limited annotations, the multi-site training data from different domains makes the task remain challenging. Due to domain shift, the existing methods cannot be well generalized to the out-of-set data, which limits the practical application of deep learning in the field of medical imaging. In this work, we propose an effective domain adaptation framework which consists of a bidirectional image translation module and two symmetrical image segmentation modules. The framework improves the generalization ability of deep neural networks in medical image segmentation. The image translation module conducts the mutual conversion between the source domain and the target domain, while the symmetrical image segmentation modules perform image segmentation tasks in both domains. Besides, we utilize adversarial constraint to further bridge the domain gap in feature space. Meanwhile, a consistency loss is also utilized to make the training process more stable and efficient. Experiments on a multi-site ultrasound thyroid nodule dataset achieve 96.22% for PA and 87.06% for DSC in average, demonstrating that our method performs competitively in cross-domain generalization ability with state-of-the-art segmentation methods.

Introduction

Deep learning models for medical imaging often suffer from poor generalization. A model trained to perfection on ultrasound images from one hospital often fails when deployed at another because of differences in imaging equipment, operator protocols, and patient demographics—a phenomenon known as domain shift. In the context of thyroid nodules, manual segmentation is labor-intensive and requires high-level expertise. While automated tools exist, they struggle with the "out-of-distribution" data found in multi-site clinical settings. This research addresses the challenge of unsupervised domain adaptation (UDA): leveraging labeled data from a source site to perform accurately on unlabeled data from a target site.

Proposed Methodology

The authors propose a complex, multi-stage architecture that aligns data at both the pixel level (visual appearance) and the feature level (latent representation).

The Symmetrical Awareness Network framework, illustrating bidirectional translation and dual segmentation branches with consistency loss.

1. Bidirectional Image Translation

Inspired by CycleGAN, this module learns to translate images from the source domain to the target domain ($G_{S \to T}$) and vice versa ($G_{T \to S}$). This allows the model to "practice" on images that mimic the visual characteristics of the target site. The generators in this module utilize an encoder-decoder structure with residual blocks to preserve spatial information during translation.

The encoder-decoder architecture used within the translation modules, featuring convolutional layers and ResNet blocks.

2. Symmetrical Segmentation Modules

Instead of a single network, the framework employs two identical segmentation branches ($F_S$ and $F_T$). These use EfficientNet-B0 as a backbone and incorporate a hybrid attention mechanism (channel and spatial attention) to better capture nodule boundaries.

3. Multi-Level Adversarial Learning and Consistency

The system uses four discriminators—two for pixels and two for features—to force the latent representations of both domains into the same distribution. A key innovation is the Segmentation Consistency Loss. The authors argue that an image and its translated counterpart (e.g., a source image and its "target-styled" version) should produce the same segmentation mask. This consistency acts as a powerful supervisor for the unlabeled target data.

Experimental Setup

The study utilized three distinct ultrasound thyroid nodule datasets to simulate cross-site scenarios:

  • Private Dataset 1: 936 images.
  • Private Dataset 2: 740 images.
  • Public Dataset (DDTI): 637 images from the Data Database Thyroid Ultrasound.

Researchers conducted six sets of experiments, rotating which dataset served as the "source" and which served as the "target."

Key Results and Insights

The proposed method consistently outperformed state-of-the-art (SOTA) segmentation models like DeepLabV3, PSPNet, and FPN in cross-site tests.

  • Performance Metrics: The framework achieved an average 96.22% Pixel Accuracy (PA) and 87.06% Dice Similarity Coefficient (DSC) across all test sets.
  • Domain Adaptation Impact: Compared to standard segmentation without domain adaptation, the method increased the average DSC by 3.64%.
  • Two-Step Alignment: Aligning only visual style (pixels) is insufficient for medical precision. By also aligning latent vectors (features), the model ensures semantic consistency across domains.
  • The Power of Symmetry: Maintaining two symmetrical branches prevents the network from "forgetting" the source domain while learning the target domain, leading to more stable training.

Conclusion and Future Work

The Symmetrical Awareness Network demonstrates that combining bidirectional translation with dual-branch consistency significantly mitigates domain shift in medical imaging. However, the authors note that the computational complexity of multiple generators and discriminators is high. Future work may explore domain generalization, where a model is trained to be robust to any new site without requiring a specific adaptation phase, and optimize for real-time inference speed during live ultrasound examinations.

Summary updated

Abstract

Recent years have seen remarkable progress of learning-based methods on Ultrasound Thyroid Nodules segmentation. However, with very limited annotations, the multi-site training data from different domains makes the task remain challenging. Due to domain shift, the existing methods cannot be well generalized to the out-of-set data, which limits the practical application of deep learning in the field of medical imaging. In this work, we propose an effective domain adaptation framework which consists of a bidirectional image translation module and two symmetrical image segmentation modules. The framework improves the generalization ability of deep neural networks in medical image segmentation. The image translation module conducts the mutual conversion between the source domain and the target domain, while the symmetrical image segmentation modules perform image segmentation tasks in both domains. Besides, we utilize adversarial constraint to further bridge the domain gap in feature space. Meanwhile, a consistency loss is also utilized to make the training process more stable and efficient. Experiments on a multi-site ultrasound thyroid nodule dataset achieve 96.22% for PA and 87.06% for DSC in average, demonstrating that our method performs competitively in cross-domain generalization ability with state-of-the-art segmentation methods.
BibTeX
@inproceedings{ma2023symmetrical,
  author = {Wenxuan Ma and Xiaopeng Li and Lian Zou and Cien Fan and Meng Wu},
  title = {{Symmetrical awareness network for cross-site ultrasound thyroid nodule segmentation}},
  booktitle = {Frontiers in Public Health},
  year = {2023},
  doi = {10.3389/fpubh.2023.1055815},
}
Figure