Skip to main content
Back to timeline
arXivSource publication:

Ultrasound super-resolution stops guessing degradations from semantic tags: DAP-MR splits type from strength and holds up across anatomy shift

Synopsis

Diffusion super-resolution models built for natural images often rely on text prompts when applied to ultrasound, but automatically generated tags do not match the actual restoration needs, and medical wording can be misread by general-purpose text encoders. DAP-MR separates degradation identity from degradation strength: each degradation gets a dedicated CLIP-initialized token, and a router predicts a continuous magnitude that modulates that token. Across four diffusion super-resolution backbones, DAP-MR raises both PSNR and SSIM on all four on the anatomy-shifted BUSI set, and raises Dice in all eight backbone-dataset comparisons on two unpaired real-world ultrasound sets.

Source-provided article image: DAP-MR: Degradation-Aware Prompt-Magnitude Routing for Real-World Ultrasound Super-Resolution
arXiv

Interpretation

Ultrasound super-resolution models can now be guided by two independent signals, which degradation is present and how strong it is, instead of packing both into one text prompt. When prompt strategies from natural-image super-resolution were moved directly to ultrasound, the degradation-aware prompt extractor produced tags describing generic visual attributes rather than the ultrasound structures or degradations to be restored, giving only limited change over the no-text baseline; medical captions also gave less predictable restoration guidance in the comparison. Across four diffusion super-resolution backbones, DAP-MR improves both PSNR and SSIM for all four on the anatomy-shifted BUSI set, while on the in-domain Cardiac set it improves 15 of 16 PSNR and SSIM comparisons, with SeeSR's PSNR decreasing slightly.

The design lets language say only which degradation to address, while a separate continuous number supplies the strength. Each degradation has a dedicated token initialized from the contextualized EOS representation of a degradation-oriented CLIP prompt and then adapted during ultrasound super-resolution training; a separate router regresses an input-dependent continuous magnitude from the frozen VAE latent and uses it to modulate the corresponding token. The router is pretrained on synthetically degraded ultrasound images with known continuous magnitude labels drawn only from the Cardiac training set; during super-resolution training, ground-truth magnitudes are used first and then switched to router predictions to match blind inference conditions.

Training adds two further supervision modes so the model actually uses the magnitude signal and the degradation identity. With only clean high-resolution targets for full-restoration supervision, the backbone has limited incentive to rely on degradation-token conditioning, since restoration can be learned largely from visual features and the backbone prior. Partial-dose supervision uses only a fraction of the input degradation magnitude as the control and keeps the remaining degradation in the target; mismatch-specificity supervision activates a token that does not match the input degradation and uses the degraded input itself as the target, suppressing cross-degradation responses.

Perspective

Evaluation covers 1,417 echocardiography frames from EchoNet with an 8:2 patient-level train-test split, 768 BUSI breast ultrasound images, and CAMUS-Poor with 232 low-quality echocardiography images and HMC-QU with 473 clinical ultrasound images, neither used in training nor synthetically degraded; paired synthetic evaluation uses PSNR, SSIM and LPIPS, and unpaired real data uses Dice from a frozen SAMUS segmentation model.

In-domain Cardiac gains are described as relatively modest and occasionally mixed, with SeeSR's PSNR decreasing slightly; LPIPS improves in only four of eight backbone-dataset comparisons, potentially reflecting backbone-dependent trade-offs and limited transfer of natural-image-pretrained perceptual features to ultrasound. The router keeps high AUROC for degradation presence on BUSI, but magnitude estimation error rises, particularly for attenuation, and predictions become more conservative and tend to underestimate stronger degradations, so absolute magnitude calibration is sensitive to domain shift. The authors also state that magnitude values are continuous conditioning variables rather than exact percentages of degradation removal, and stronger control does not always mean better restoration.