An Uncertainty-guided Tiered Self-training Framework for Active Source-free Domain Adaptation in Prostate Segmentation

TL;DR

This work tackles cross-center prostate MRI segmentation by proposing an active source-free domain adaptation (ASFDA) framework, UGTST, that leverages a single round of target-domain inference. It introduces Global Aleatoric Uncertainty Aggregation (GAUA) to quantify global uncertainty from test-time augmented predictions and a diversity-aware filter to select a representative, informative annotated subset within a small budget. A tiered self-training strategy then uses labeled active samples and progressively refined pseudo-labels to achieve robust target-domain adaptation. Across two target centers, UGTST substantially outperforms prior SFDA methods and rivals fully supervised performance with only 5% annotation, highlighting its clinical practicality for privacy-preserving domain adaptation in medical imaging.

Abstract

Deep learning models have exhibited remarkable efficacy in accurately delineating the prostate for diagnosis and treatment of prostate diseases, but challenges persist in achieving robust generalization across different medical centers. Source-free Domain Adaptation (SFDA) is a promising technique to adapt deep segmentation models to address privacy and security concerns while reducing domain shifts between source and target domains. However, recent literature indicates that the performance of SFDA remains far from satisfactory due to unpredictable domain gaps. Annotating a few target domain samples is acceptable, as it can lead to significant performance improvement with a low annotation cost. Nevertheless, due to extremely limited annotation budgets, careful consideration is needed in selecting samples for annotation. Inspired by this, our goal is to develop Active Source-free Domain Adaptation (ASFDA) for medical image segmentation. Specifically, we propose a novel Uncertainty-guided Tiered Self-training (UGTST) framework, consisting of efficient active sample selection via entropy-based primary local peak filtering to aggregate global uncertainty and diversity-aware redundancy filter, coupled with a tiered self-learning strategy, achieves stable domain adaptation. Experimental results on cross-center prostate MRI segmentation datasets revealed that our method yielded marked advancements, with a mere 5% annotation, exhibiting an average Dice score enhancement of 9.78% and 7.58% in two target domains compared with state-of-the-art methods, on par with fully supervised learning. Code is available at:https://github.com/HiLab-git/UGTST

Figures (3)

• Figure 1: Overview of our Uncertainty-guided Tiered Self-training Framework, where the $D_t$, $D_{tu}$, $D_{ts}$ and $D_{ta}$ are the target domain set, uncertainty candidate set, assumed stable set, and active sample set, respectively. Our method uses the augmentation-based perturbations output for active sample selection via uncertainty and diversity, then employs a tiered self-training strategy for domain adaptation.
• Figure 2: Qualitative comparison of different Domain Adaptation methods. The ground truth and prediction are displayed in yellow and green contours respectively.
• Figure 3: Ablation study on the validation set. (a). The effect of capacity of $D_{tu}$ with $M=5\%$, (b). Comparison between different uncertainty estimation methods with two-stage results, and (c). Comparison of semi-supervised learning and our method.