Abstract
Unsupervised domain adaptation (UDA) is an emerging technique that enables the transfer of domain knowledge learned from a labeled source domain to unlabeled target domains, providing a way of coping with the difficulty of labeling in new domains. The majority of prior work has relied on both source and target domain data for adaptation. However, because of privacy concerns about potential leaks in sensitive information contained in patient data, it is often challenging to share the data and labels in the source domain and trained model parameters in cross-center collaborations. To address this issue, we propose a practical framework for UDA with a black-box segmentation model trained in the source domain only, without relying on source data or a white-box source model in which the network parameters are accessible. In particular, we propose a knowledge distillation scheme to gradually learn target-specific representations. Additionally, we regularize the confidence of the labels in the target domain via unsupervised entropy minimization, leading to performance gain over UDA without entropy minimization. We extensively validated our framework on a few datasets and deep learning backbones, demonstrating the potential for our framework to be applied in challenging yet realistic clinical settings.
Introduction
Deep learning models for medical imaging often suffer from significant performance degradation when deployed in new environments, a phenomenon known as domain shift. A model optimized for MRI scans at one facility may fail at another due to variations in scanner hardware, imaging protocols, or patient demographics.
Standard Unsupervised Domain Adaptation (UDA) techniques typically require access to either the original source training data or the internal weights of the source model ("white-box" access). In clinical practice, strict privacy regulations like HIPAA often prohibit data sharing. Furthermore, sharing model weights can expose sensitive information through "deep inversion" attacks and mandates that the target site use the identical network architecture. This paper introduces a Black-Box UDA (BBUDA) framework that requires only API-level access to the source model, preserving privacy and offering architectural flexibility.
Methodology
The proposed framework utilizes a teacher-student paradigm where a black-box source model serves as the teacher to guide a student model in the target domain.
Teacher-Student Distillation with EMD
Because of the domain shift, the teacher’s predictions on target data are inherently noisy. To mitigate error propagation, the authors introduce Exponential Mixup Decay (EMD). This strategy generates pseudo-labels by blending the teacher’s prediction with the student’s current prediction. Early in training, the student relies heavily on the teacher's knowledge. As training progresses, an exponential decay function reduces this reliance, encouraging the student to "self-train" and specialize in the target domain's unique features.
Entropy Minimization
To further refine the student's performance, the framework incorporates unsupervised entropy minimization. This loss function penalizes uncertain, high-entropy predictions, forcing the model to make more confident, pixel-wise classifications. This dual approach—distillation for initialization and entropy minimization for refinement—allows the student to surpass the noisy guidance of the source model.
Experimental Setup
The framework was validated using the BraTS 2018 database, focusing on multi-modal MRI scans (T1, T1ce, T2, FLAIR). The researchers tested cross-subtype adaptation (e.g., High-Grade Glioma to Low-Grade Glioma) and cross-modality adaptation (e.g., T2 to T1-weighted scans). Notably, the framework is model-agnostic, allowing a complex U-Net teacher to train a lightweight MobileNet student.
Results and Discussion
The BBUDA method was compared against "Source Only" baselines and "OSUDA," a state-of-the-art white-box method.
Quantitative Performance
In the HGG to LGG adaptation task, BBUDA significantly improved the Dice Score for the "Whole Tumor" from 79.29% to 82.21%. This performance closely approached the 83.62% achieved by the white-box OSUDA method, which requires full parameter access. Furthermore, the Hausdorff Distance—a measure of boundary accuracy—dropped from 38.7mm to 28.6mm, indicating much sharper and more accurate tumor edge detection.
Key Insights
The success of BBUDA demonstrates that privacy constraints do not have to be a barrier to effective model adaptation. The core innovation is the Exponential Mixup Decay; by acknowledging that the teacher is only useful for initial guidance, the student is empowered to eventually trust its own learned representations of the target data. This flexibility also enables the distillation of knowledge from heavy server-side models into efficient models suitable for edge deployment in clinics.
Conclusion
This work provides a robust, privacy-preserving pathway for deploying medical AI across different clinical centers. While currently focused on 2D architectures, future extensions to 3D volumes could further enhance spatial consistency. Additionally, developing automated methods to balance distillation and entropy losses will be crucial for making these systems "plug-and-play" for clinical practitioners.
Summary updated