Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Federated Self-supervised Domain Generalization for Label-efficient Polyp Segmentation

Xinyi Tan, Jiacheng Wang, Liansheng Wang

TL;DR

The paper tackles label-efficient polyp segmentation under privacy constraints by proposing LFDG, a Federated self-supervised Domain Generalization framework. It combines DropPos-based self-supervision with self-supervised adversarial data augmentation (SSADA) and a Source-reconstruction and Augmentation-masking (SRAM) relaxation to enhance cross-domain generalization, with FedAvg used for server aggregation and a frozen backbone during downstream fine-tuning. On PolypGen data from six medical centers, LFDG achieves a Mean IoU of $0.628$ on the primary domain and $0.559$ on an unseen domain, outperforming baselines such as FedBYOL, FedMoco, and DropPos-based approaches by several percentage points, and ablations show SSADA and SRAM contribute a $\approx$2.50% and $\approx$1.30% gain respectively. The work demonstrates a privacy-preserving, data-efficient strategy that improves generalization for polyp segmentation across diverse clinical domains, with potential impact on clinical workflows where labeled data are scarce.

Abstract

Employing self-supervised learning (SSL) methodologies assumes par-amount significance in handling unlabeled polyp datasets when building deep learning-based automatic polyp segmentation models. However, the intricate privacy dynamics surrounding medical data often preclude seamless data sharing among disparate medical centers. Federated learning (FL) emerges as a formidable solution to this privacy conundrum, yet within the realm of FL, optimizing model generalization stands as a pressing imperative. Robust generalization capabilities are imperative to ensure the model's efficacy across diverse geographical domains post-training on localized client datasets. In this paper, a Federated self-supervised Domain Generalization method is proposed to enhance the generalization capacity of federated and Label-efficient intestinal polyp segmentation, named LFDG. Based on a classical SSL method, DropPos, LFDG proposes an adversarial learning-based data augmentation method (SSADA) to enhance the data diversity. LFDG further proposes a relaxation module based on Source-reconstruction and Augmentation-masking (SRAM) to maintain stability in feature learning. We have validated LFDG on polyp images from six medical centers. The performance of our method achieves 3.80% and 3.92% better than the baseline and other recent FL methods and SSL methods, respectively.

Federated Self-supervised Domain Generalization for Label-efficient Polyp Segmentation

TL;DR

The paper tackles label-efficient polyp segmentation under privacy constraints by proposing LFDG, a Federated self-supervised Domain Generalization framework. It combines DropPos-based self-supervision with self-supervised adversarial data augmentation (SSADA) and a Source-reconstruction and Augmentation-masking (SRAM) relaxation to enhance cross-domain generalization, with FedAvg used for server aggregation and a frozen backbone during downstream fine-tuning. On PolypGen data from six medical centers, LFDG achieves a Mean IoU of on the primary domain and on an unseen domain, outperforming baselines such as FedBYOL, FedMoco, and DropPos-based approaches by several percentage points, and ablations show SSADA and SRAM contribute a 2.50% and 1.30% gain respectively. The work demonstrates a privacy-preserving, data-efficient strategy that improves generalization for polyp segmentation across diverse clinical domains, with potential impact on clinical workflows where labeled data are scarce.

Abstract

Employing self-supervised learning (SSL) methodologies assumes par-amount significance in handling unlabeled polyp datasets when building deep learning-based automatic polyp segmentation models. However, the intricate privacy dynamics surrounding medical data often preclude seamless data sharing among disparate medical centers. Federated learning (FL) emerges as a formidable solution to this privacy conundrum, yet within the realm of FL, optimizing model generalization stands as a pressing imperative. Robust generalization capabilities are imperative to ensure the model's efficacy across diverse geographical domains post-training on localized client datasets. In this paper, a Federated self-supervised Domain Generalization method is proposed to enhance the generalization capacity of federated and Label-efficient intestinal polyp segmentation, named LFDG. Based on a classical SSL method, DropPos, LFDG proposes an adversarial learning-based data augmentation method (SSADA) to enhance the data diversity. LFDG further proposes a relaxation module based on Source-reconstruction and Augmentation-masking (SRAM) to maintain stability in feature learning. We have validated LFDG on polyp images from six medical centers. The performance of our method achieves 3.80% and 3.92% better than the baseline and other recent FL methods and SSL methods, respectively.

Paper Structure

This paper contains 15 sections, 8 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Overall Framework
  4. Self-supervised Adversarial Data Augmentation
  5. Maximization Phase
  6. Minimization Phase
  7. Source-reconstruction and Augmentation-masking
  8. Loss Function
  9. Experiments and Results
  10. Dataset and Implementation Details
  11. Dataset:
  12. Metrics and Implementation Details:
  13. Comparison with Prior Work
  14. Ablation Studies
  15. Conclusion

Figures (2)

  • Figure 1: The overall framework of LFDG. It learns generalizable representations through self-supervised adversarial data augmentation (SSADA) and source-reconstruction and Augmentation-masking(SRAM) during the local training step and updates the learned parameters in the server.
  • Figure 2: Polyp segmentation visualization on the PolypGen dataset.