[CVPR 2025] Tell2Adapt: Bridging the "Extreme Gap" in Medical SFUDA with Vision Foundation Models

TL;DR

Source-Free Unsupervised Domain Adaptation (SFUDA) is a "holy grail" for clinical AI deployment, where data privacy prevents access to original training data. Tell2Adapt breaks the limitations of existing specialized methods by using a Vision Foundation Model (VFM) guided by LLM-refined text prompts. It successfully bridges extreme domain gaps (like MR to Ultrasound) across 22 anatomical targets, achieving SOTA results while providing a lightweight student model for clinical use.

Problem & Motivation: The Failure of Spatial Prompts

Traditional SFUDA methods often rely on the source model to generate pseudo-labels or spatial prompts (bounding boxes/points) for refinement. However, when the domain gap is "huge"—such as migrating a model from high-resolution MRI to noisy Ultrasound—the source model's predictions are often "garbage."

Using these failed predictions as spatial prompts for foundation models like SAM (Segment Anything Model) leads to a failure cascade: noisy prompts lead to noisy pseudo-labels, which ultimately lead to model collapse. The researchers identified that while visual domains shift drastically, semantic instructions (e.g., "Spleen in abdominal CT") remain constant.

Methodology: Tell, then Adapt

Tell2Adapt introduces two key innovations to stabilize the adaptation process:

1. Context-Aware Prompts Regularization (CAPR)

Instead of relying on imperfect spatial cues, Tell2Adapt uses text. To prevent "semantic drift" caused by noisy or ambiguous user queries, an LLM (Qwen-series) acts as a semantic normalizer. It analyzes all input queries, infers global context (modality and site), and generates a canonical format: [Target] in [Anatomical Site] [Modality].

2. Visual Plausibility Refinement (VPR)

Even after distillation, student models may produce "anatomically impossible" artifacts. VPR acts as a clinical gatekeeper. It leverages pre-computed Beta distributions from the VFM that describe the visual attributes (color, intensity, probability) of specific organs. Any predicted component that falls outside the statistical "norm" is pruned.

Experiments: Breaking the Extreme Domain Gap

The authors conducted one of the most extensive evaluations in SFUDA history, covering 10 adaptation directions.

• The Impossible Task: In MR to Ultrasound (MR→US) cardiac segmentation—a task where most traditional SFUDA methods fail completely (DICE near 0)—Tell2Adapt achieved 94.6% DICE for the Left Ventricle.
• Efficiency: While the VFM (BiomedParse) requires a massive 26.1 GB of VRAM, the adapted student model uses only 4.8 GB, making it deployable on consumer-grade medical hardware.

Critical Analysis & Conclusion

The genius of Tell2Adapt lies in decoupling. By moving away from spatial prompts and toward text-guided distillation, the framework bypasses the "source model reliability" trap.

Takeaway: This work demonstrates that the future of domain adaptation isn't just about better alignment algorithms, but about effectively "interrogating" the vast world knowledge locked within Vision-Language models.

Limitations: While the final model is fast, the pseudo-label generation phase still requires significant compute time (up to 3.4s per volume), which might hinder real-time "on-the-fly" adaptation in emergency settings.

For those interested in the implementation, the authors have released their canonical meta-prompts and ablation code, providing a new benchmark for robust prompt engineering in medical AI.