TL;DR
DAP-MR (Degradation-Aware Prompt–Magnitude Routing) improves ultrasound super-resolution by decoupling the type of image degradation from its severity. By using text-derived tokens for degradation identity and a learned "magnitude router" for continuous strength control, the framework achieves superior restoration quality and generalizes far better to unseen anatomical domains than standard text-conditioned models.
What Problem Does This Paper Address?
Super-resolution (SR) for ultrasound is notoriously difficult due to complex, scanner-specific degradations like speckle noise, signal attenuation, and depth-dependent blur. While modern diffusion-based SR models (like Stable Diffusion) offer powerful generative priors, they rely on text prompts that don't translate well to the medical domain.
The authors identify three main hurdles:
- Semantic Mismatch: Natural image tags (e.g., "high resolution") provide little guidance for ultrasound structures.
- Medical Jargon: General-purpose text encoders (like CLIP) often fail to grasp specialized medical terminology.
- Entanglement: Existing methods that use prompts like "strong blur" or "blur level 0.8" force the model to learn both the identity of the defect and its severity from the same embedding. This often leads to ambiguous restoration, resulting in over-smoothing or unwanted artifacts.
Methods
The core innovation of DAP-MR is factorized conditioning. Instead of a single text prompt, the system breaks degradation down into two components:
- Degradation Tokens (The "What"): The authors define $K=6$ common ultrasound degradations (blur, resize, attenuation, speckle, sensor noise, and anechoic haze). Each is represented by a dedicated token initialized from the CLIP embedding of a descriptive prompt (e.g., "speckle noise"). Crucially, these tokens remain learnable during training to adapt to ultrasound-specific appearance statistics.
- Magnitude Router (The "How Much"): A separate, lightweight MLP router analyzes the input image to predict a continuous magnitude vector $\hat{m} \in [0, 1]^K$. This vector modulates the degradation tokens before they enter the SR backbone's cross-attention layers.
To ensure the model actually listens to these controls, the authors introduce a multi-mode training strategy:
- Full-restoration: Standard mapping from degraded input to clean target.
- Partial-dose Supervision: The model is trained to reach an intermediate target based on a fraction of the input degradation. This teaches the model that a magnitude of 0.5 means "fix half the blur."
- Mismatch-specificity: The model is given a token for a degradation that isn't present (e.g., a "noise" token for a "blur" image) and is penalized if it changes the image. This prevents different degradation controls from interfering with each other.
Datasets
The models were trained and evaluated on a mix of synthetic and real-world data:
- EchoNet (Cardiac): Used as the primary in-domain training and testing set.
- BUSI (Breast Ultrasound): Used to test "anatomical-domain shift"—evaluating if a model trained on hearts can generalize to breast tissue.
- CAMUS-Poor & HMC-QU: Unpaired, real-world clinical datasets used for downstream evaluation.
Key Insights and Ideas
- Language as a Semantic Anchor: Using CLIP to initialize tokens provides a meaningful starting point, but letting those tokens evolve during training is what allows the model to bridge the gap between natural language and medical imaging.
- Continuous Control is King: Unlike discrete text prompts ("low," "medium," "high"), the magnitude router provides a continuous signal. This allows for much finer restoration, especially in real-world cases where degradation levels are unpredictable.
- Generalization through Degradation: By focusing the conditioning on how the image is broken rather than what the organ is, the model becomes much more robust to anatomical shifts. A "blur" token is relevant whether you are looking at a heart or a liver.
Key Results
DAP-MR was integrated into four different diffusion backbones (SeeSR, StableSR, PiSA-SR, and TVT) and consistently outperformed the baselines:
- Out-of-Domain Performance: On the BUSI dataset (breast ultrasound), PiSA-SR with DAP-MR saw a massive +6.56% boost in PSNR and a +25.39% improvement in SSIM compared to the baseline.
- Clinical Utility: The authors tested the SR results using a frozen downstream segmentation model (SAMUS). DAP-MR improved Dice scores across the board, notably increasing the score on the HMC-QU clinical dataset by up to 18.77% for the TVT backbone.
- Qualitative Sharpness: As seen in Figure 4, DAP-MR produces clearer boundaries and more realistic textures compared to standard prompting methods, which often leave residual blur or create "plastic-looking" artifacts.
Limitations and Future Work
While the router generalizes well in terms of detecting degradations, the authors noted that absolute magnitude calibration remains sensitive to domain shifts. For example, the router might underestimate the severity of blur when moving from cardiac to breast ultrasound.
Future work could explore more robust, self-supervised ways to calibrate these magnitudes or investigate if this factorized approach can be extended to other medical modalities like MRI or CT where similar "entanglement" issues exist.
Summary updated