Generalizable Pancreas Segmentation via a Dual Self-Supervised Learning Framework

Jun Li, Hongzhang Zhu, Tao Chen, Xiaohua QianIEEE JBHI 2023 · DOI · OpenReview

TL;DR

This paper tackles the critical challenge of pancreas segmentation models failing when applied to data from different hospitals or scanners (generalization). They introduce a novel dual self-supervised learning framework that leverages both global anatomical context and local image restoration to significantly improve robustness and accuracy on unseen data, outperforming existing methods.

What Problem Does This Paper Address?

Medical image segmentation models, particularly those trained on Convolutional Neural Networks (CNNs), often perform exceptionally well on the specific datasets they were trained on. However, this performance dramatically drops when these models are applied to data from different sources (e.g., different hospitals, scanners, or imaging protocols). This phenomenon, known as the "generalization gap," is a major hurdle for deploying AI in real-world clinical settings. The pancreas, in particular, is a challenging organ to segment due to its variable shape, obscure boundaries, and location amidst other abdominal organs. Existing methods often fail to account for these variations, leading to unreliable segmentation results when applied to new datasets. The authors aim to improve the single-source generalization of pancreas segmentation models, meaning a model trained on data from one source should perform well on data from other sources without requiring retraining or fine-tuning on the new data.

Methods

The core of this paper is a novel dual self-supervised learning framework designed to enhance pancreas segmentation generalization. It operates in a coarse-to-fine manner, with the proposed method being applied in the fine segmentation stage. The framework comprises two key modules:

  1. Global-Feature Self-Supervised Contrastive Learning Module: This module aims to create more robust and discriminative feature representations. It achieves this by:
  2. Maximizing intra-class cohesion: Ensuring that features from different parts of the pancreas are clustered together. This is done by sampling "positive" features along the pancreas's centerline, capturing its spatial continuity.
  3. Maximizing inter-class separation: Clearly distinguishing pancreas features from surrounding non-pancreatic tissues. This is achieved by using contrastive learning to push apart features from the pancreas and its background. This helps mitigate the influence of surrounding tissues, which can be a major source of error in high-uncertainty regions.
  4. Local-Image-Restoration Self-Supervised Learning Module: This module focuses on improving the characterization of "high-uncertainty regions" – areas where segmentation is typically difficult due to blurred boundaries or poor anatomical context. It works by:
  5. Identifying high-uncertainty regions: Using pixel-level uncertainty maps derived from multiple augmented views of the input image.
  6. Corrupting and restoring appearance: Randomly corrupting the appearance of these high-uncertainty regions in the input image and then training a module to restore them. This restoration process leverages the local anatomical context, forcing the model to learn fine-grained details and spatial relationships.
  7. Synergistic learning: This module is trained alongside the segmentation task, using both restoration and segmentation losses.

The framework uses a shared encoder for both modules, promoting efficient learning.

Datasets

The authors evaluated their method on three distinct pancreas CT imaging datasets:

  • NIH Dataset: A public dataset from the National Institutes of Health Clinical Center, containing normal subjects.
  • MSD Dataset: A public dataset from the Medical Segmentation Decathlon, collected from patients undergoing resection of pancreatic masses (including IPMN, PNET, and PDA). This dataset contains artifacts like biliary stents.
  • Private Dataset: A dataset collected from Renji Hospital of Shanghai Jiao Tong University School of Medicine, specifically comprising patients with pancreatic ductal adenocarcinoma (PDA).

These datasets were used for both training and testing, with a focus on evaluating the model's performance on unseen data from different sources.

Key Insights and Ideas

  • High-uncertainty regions are key to generalization: The paper strongly posits that performance degradation in generalization tasks is primarily driven by difficulties in segmenting these high-uncertainty regions. Focusing on improving their characterization is crucial.
  • Global and local context are complementary: The dual approach, combining global feature separation with local image restoration, is more effective than either method alone. The global module handles broader anatomical distinctions, while the local module refines details in challenging areas.
  • Anatomical context is a powerful prior: The framework effectively leverages the inherent spatial continuity and structure of the pancreas, using it as a strong prior to guide segmentation, especially in ambiguous areas.
  • Self-supervision is vital for limited data: The use of self-supervised learning allows the model to learn rich representations from unlabeled or partially labeled data, which is essential when labeled multi-source data is scarce.

Key Results

The proposed dual self-supervised learning framework achieved state-of-the-art performance on the single-source generalization task.

  • Improved Generalization: The method demonstrated significant performance gains across all three datasets when tested on unseen sources. For example, on the Private dataset (trained on NIH), the proposed method achieved a DSC of 70.08%, outperforming the baseline (3D-C2F) which achieved 69.76% in the paper's ablation study (Table IV).
  • Reduced Uncertainty: The method substantially reduced the number of pixels with high uncertainty (Table V), indicating more confident and stable segmentations.
  • Ablation Studies: Comprehensive ablation studies confirmed the effectiveness of each component. Adding the Global-Feature Self-Supervised (GFS) module and the Local-Image-Restoration Self-Supervised (LIS) module individually to a baseline (3D-C2F) showed incremental improvements, but their combination (Ours) yielded the best results.
  • Computational Efficiency: The proposed method achieved competitive performance with a relatively low number of parameters and reasonable inference time, making it practical for clinical deployment.

Limitations and Future Work

  • Data Disparity: The authors acknowledge that significant data disparities (e.g., presence of biliary stents, different tumor types, normal vs. pathological subjects) can still pose challenges, particularly when training on one dataset and testing on a vastly different one. For instance, the LIS module trained on MSD performed poorly on NIH due to the absence of stents and heterogeneous lesions.
  • Future Work: The authors suggest exploring how to better facilitate information transfer and fusion between the local-image-restoration and segmentation tasks. They also aim to address challenges related to samples exhibiting different health conditions and the curation of multi-pathology pancreatic imaging data to further mitigate the impact of heterogeneous patient populations on generalization performance.

Summary updated

Abstract

Recently, numerous pancreas segmentation methods have achieved promising performance on local single-source datasets. However, these methods don't adequately account for generalizability issues, and hence typically show limited performance and low stability on test data from other sources. Considering the limited availability of distinct data sources, we seek to improve the generalization performance of a pancreas segmentation model trained with a single-source dataset, i.e., the single source generalization task. In particular, we propose a dual self-supervised learning model that incorporates both global and local anatomical contexts. Our model aims to fully exploit the anatomical features of the intra-pancreatic and extra-pancreatic regions, and hence enhance the characterization of the high-uncertainty regions for more robust generalization. Specifically, we first construct a global-feature contrastive self-supervised learning module that is guided by the pancreatic spatial structure. This module obtains complete and consistent pancreatic features through promoting intra-class cohesion, and also extracts more discriminative features for differentiating between pancreatic and non-pancreatic tissues through maximizing inter-class separation. It mitigates the influence of surrounding tissue on the segmentation outcomes in high-uncertainty regions. Subsequently, a local-image restoration self-supervised learning module is introduced to further enhance the characterization of the high uncertainty regions. In this module, informative anatomical contexts are actually learned to recover randomly corrupted appearance patterns in those regions.
BibTeX
@article{li2023generalizable,
  author = {Jun Li and Hongzhang Zhu and Tao Chen and Xiaohua Qian},
  title = {{Generalizable Pancreas Segmentation via a Dual Self-Supervised Learning Framework}},
  journal = {IEEE J. Biomed. Health Informatics},
  year = {2023},
  doi = {10.1109/jbhi.2023.3294278},
}
Figure