Cross-Organ Domain Adaptive Neural Network for Pancreatic Endoscopic Ultrasound Image Segmentation

ZhiChao Yan, Hui Xue, Yi Zhu, Bin Xiao, Hao YuanarXiv 2024 · DOI · arXiv

Abstract

Researchers have developed COTS-Nets, a novel deep learning framework that improves pancreatic tumor segmentation in endoscopic ultrasound (EUS) images by leveraging knowledge from other organs, such as the thyroid and ovaries. By mimicking the cross-organ diagnostic patterns used by radiologists, this cross-organ domain adaptation approach achieves state-of-the-art accuracy even with limited and low-quality pancreatic data.

Introduction: The Challenge of Pancreatic EUS

Pancreatic cancer remains one of the most lethal malignancies, with a five-year survival rate of less than 7%. Early diagnosis is critical, and Endoscopic Ultrasound (EUS) is a primary tool for local detection. However, developing AI for EUS segmentation faces two significant hurdles:

  1. Data Scarcity: High-quality, labeled EUS images of the pancreas are difficult to acquire.
  2. Image Quality: EUS images are inherently noisy, low-contrast, and often feature indistinct boundaries between tumors and healthy tissue.
Clinical context of pancreatic Endoscopic Ultrasound (EUS) and the challenge of identifying tumor masses within complex anatomy.

While traditional domain adaptation (DA) methods bridge gaps between different views of the same organ (e.g., CT to MRI), this research introduces Cross-Organ Domain Adaptation. It explores whether ultrasound data from other organs—where tumors share similar textures or "echo" patterns—can enhance a model's understanding of the pancreas.

Methodology: COTS-Nets

The proposed Cross-Organ Tumor Segmentation Networks (COTS-Nets) comprise two primary components working in tandem:

The Universal Network

This is the core segmentation engine, based on a U-shaped Vision Transformer. It incorporates a specialized Boundary Loss—utilizing Sobel and Gaussian filters—to emphasize the fuzzy edges of tumors. Additionally, a Consistency Loss based on an Exponential Moving Average (EMA) aligns features learned from different organs, narrowing the "domain gap."

The Auxiliary Network

Serving as a "knowledge bridge," this network uses a Hybrid Adaptive Attention Module (HAAM) to extract multi-scale spatial and channel features from both the source organ (e.g., thyroid) and the target organ (pancreas). This enables the Universal Network to learn "domain-invariant" features—characteristics of a tumor that remain consistent regardless of the organ.

The training objective is a multi-task loss function combining standard segmentation loss with distillation, boundary, and consistency losses.

Experimental Setup and Datasets

The study evaluates the framework using three primary ultrasound datasets:

  1. MMOTU (Ovarian): 1,469 2D ultrasound images used as a source domain.
  2. TN3K (Thyroid): 3,493 grayscale ultrasound images used as another source domain.
  3. PCEUS (Pancreas): A newly introduced dataset containing 501 pathologically confirmed pancreatic EUS images, providing a high-quality benchmark for this data-poor field.

Key Insights

  • Radiologist-Inspired Learning: The core philosophy mimics medical training, where radiologists learn "meta-knowledge" of lesion characteristics (texture, echoes, boundary behavior) across various regions before specializing.
  • Morphological Consistency: The authors exploit the fact that tumors across different organs often exhibit similar microscopic morphological traits, resulting in comparable visual patterns in ultrasound.
  • Boundary-First Approach: By prioritizing boundary detection as a primary loss constraint, the model achieves significantly sharper segmentations in low-contrast environments.

Results

COTS-Nets outperformed several state-of-the-art models, including U-Net, Swin-Unet, and MDViT.

  • Accuracy: In the "Thyroid → Pancreas" task, COTS-Nets achieved a Dice score of 81.65% and an IoU of 69.50%.
  • Boundary Precision: The Average Symmetric Distance (ASD) dropped to 0.42mm, a significant improvement over the 0.62mm achieved by the next best model.
  • Ablation Studies: Removing the Auxiliary Network or the Boundary Loss led to measurable performance declines, validating the importance of cross-organ feature fusion.

Conclusion and Future Work

While effective, the inclusion of cross-organ data can occasionally introduce slight biases, as seen in fluctuations in the Hausdorff Distance (HD95) during some tests. Future research will focus on testing the framework with a broader variety of source organs and moving toward real-time clinical integration to assist gastroenterologists during live EUS procedures.

Summary updated

BibTeX
@misc{yan2024cross,
  author = {ZhiChao Yan and Hui Xue and Yi Zhu and Bin Xiao and Hao Yuan},
  title = {{Cross-Organ Domain Adaptive Neural Network for Pancreatic Endoscopic Ultrasound Image Segmentation}},
  year = {2024},
  eprint = {2409.04718},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.04718},
}
Figure