Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DDFP: Data-dependent Frequency Prompt for Source Free Domain Adaptation of Medical Image Segmentation

Siqi Yin, Shaolei Liu, Manning Wang

TL;DR

DDFP tackles source-free domain adaptation for medical image segmentation by introducing a three-pronged approach: BN preadaptation to provide a better initialization and pseudo-labels, data-dependent frequency prompts to translate target images toward source-like appearance, and a style-focused fine-tuning regime that updates only shallow layers and prompt parameters. This end-to-end framework reduces domain gaps across the initial and subsequent training phases and leverages pseudo-labeling for self-supervision without requiring source data. Across cross-modality abdominal and cardiac datasets, DDFP achieves state-of-the-art Dice scores and competitive ASD improvements, demonstrating strong transferability with fewer trainable parameters. The work offers practical impact for privacy-conscious medical applications by enabling effective SFDA in complex segmentation tasks while remaining computationally efficient and robust to hyperparameter variations.

Abstract

Domain adaptation addresses the challenge of model performance degradation caused by domain gaps. In the typical setup for unsupervised domain adaptation, labeled data from a source domain and unlabeled data from a target domain are used to train a target model. However, access to labeled source domain data, particularly in medical datasets, can be restricted due to privacy policies. As a result, research has increasingly shifted to source-free domain adaptation (SFDA), which requires only a pretrained model from the source domain and unlabeled data from the target domain data for adaptation. Existing SFDA methods often rely on domain-specific image style translation and self-supervision techniques to bridge the domain gap and train the target domain model. However, the quality of domain-specific style-translated images and pseudo-labels produced by these methods still leaves room for improvement. Moreover, training the entire model during adaptation can be inefficient under limited supervision. In this paper, we propose a novel SFDA framework to address these challenges. Specifically, to effectively mitigate the impact of domain gap in the initial training phase, we introduce preadaptation to generate a preadapted model, which serves as an initialization of target model and allows for the generation of high-quality enhanced pseudo-labels without introducing extra parameters. Additionally, we propose a data-dependent frequency prompt to more effectively translate target domain images into a source-like style. To further enhance adaptation, we employ a style-related layer fine-tuning strategy, specifically designed for SFDA, to train the target model using the prompted target domain images and pseudo-labels. Extensive experiments on cross-modality abdominal and cardiac SFDA segmentation tasks demonstrate that our proposed method outperforms existing state-of-the-art methods.

DDFP: Data-dependent Frequency Prompt for Source Free Domain Adaptation of Medical Image Segmentation

TL;DR

DDFP tackles source-free domain adaptation for medical image segmentation by introducing a three-pronged approach: BN preadaptation to provide a better initialization and pseudo-labels, data-dependent frequency prompts to translate target images toward source-like appearance, and a style-focused fine-tuning regime that updates only shallow layers and prompt parameters. This end-to-end framework reduces domain gaps across the initial and subsequent training phases and leverages pseudo-labeling for self-supervision without requiring source data. Across cross-modality abdominal and cardiac datasets, DDFP achieves state-of-the-art Dice scores and competitive ASD improvements, demonstrating strong transferability with fewer trainable parameters. The work offers practical impact for privacy-conscious medical applications by enabling effective SFDA in complex segmentation tasks while remaining computationally efficient and robust to hyperparameter variations.

Abstract

Domain adaptation addresses the challenge of model performance degradation caused by domain gaps. In the typical setup for unsupervised domain adaptation, labeled data from a source domain and unlabeled data from a target domain are used to train a target model. However, access to labeled source domain data, particularly in medical datasets, can be restricted due to privacy policies. As a result, research has increasingly shifted to source-free domain adaptation (SFDA), which requires only a pretrained model from the source domain and unlabeled data from the target domain data for adaptation. Existing SFDA methods often rely on domain-specific image style translation and self-supervision techniques to bridge the domain gap and train the target domain model. However, the quality of domain-specific style-translated images and pseudo-labels produced by these methods still leaves room for improvement. Moreover, training the entire model during adaptation can be inefficient under limited supervision. In this paper, we propose a novel SFDA framework to address these challenges. Specifically, to effectively mitigate the impact of domain gap in the initial training phase, we introduce preadaptation to generate a preadapted model, which serves as an initialization of target model and allows for the generation of high-quality enhanced pseudo-labels without introducing extra parameters. Additionally, we propose a data-dependent frequency prompt to more effectively translate target domain images into a source-like style. To further enhance adaptation, we employ a style-related layer fine-tuning strategy, specifically designed for SFDA, to train the target model using the prompted target domain images and pseudo-labels. Extensive experiments on cross-modality abdominal and cardiac SFDA segmentation tasks demonstrate that our proposed method outperforms existing state-of-the-art methods.
Paper Structure (29 sections, 13 equations, 9 figures, 7 tables)

This paper contains 29 sections, 13 equations, 9 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Source free domain adaptation
  4. Prompt learning
  5. BN statistic calibration
  6. Methodology
  7. Problem definition
  8. Overall framework
  9. BN pre-adaptation
  10. Data-dependent frequency prompt
  11. Pseudo labeling
  12. Target model fine-tuning and loss function
  13. Experiments and results
  14. Datasets and experimental setup
  15. Datasets
  16. ...and 14 more sections

Figures (9)

  • Figure 1: Comparison between (a) the previous source-free domain adaptation framework and (b) the proposed DDFP framework. The DDFP framework reduces the domain gap throughout both the initial and subsequent training phases by utilizing a preadapted model, data-dependent frequency prompt learning, and pseudo-labeling strategies.
  • Figure 2: Overview of the proposed DDFP architecture. We introduce a BN preadaptation strategy (yellow) to initialize the target model and generate high-quality pseudo-labels for target domain data (blue). The data-dependent frequency prompt learning strategy (red) facilitates image style translation. Both the data-dependent frequency prompt parameters and the style-related layers of the target model are jointly trained to achieve DA.
  • Figure 3: Data-dependent frequency prompt generation process for each image in the target domain batch.
  • Figure 4: Visualization of SFDA segmentation results on the multiorgan abdominal dataset. The first two rows show the results for CT to MRI adaptation, while the last two rows display results for MRI to CT adaptation.
  • Figure 5: Visualization of SFDA segmentation results on the cardiac dataset. The first two rows correspond to CT to MRI adaptation, while the last two rows correspond to MRI to CT adaptation.
  • ...and 4 more figures