Blind Ultrasound Image Enhancement via Self-Supervised Physics-Guided Degradation Modeling

Ultrasound (US) interpretation is hampered by multiplicative speckle, acquisition blur from the point-spread function (PSF), and scanner- and operator-dependent artifacts. Supervised enhancement methods assume access to clean targets or known degradations; conditions rarely met in practice. We present a blind, self-supervised enhancement framework that jointly deconvolves and denoises B-mode images using a Swin Convolutional U-Net trained with a \emph{physics-guided} degradation model. From each training frame, we extract rotated/cropped patches and synthesize inputs by (i) convolving with a Gaussian PSF surrogate and (ii) injecting noise via either spatial additive Gaussian noise or complex Fourier-domain perturbations that emulate phase/magnitude distortions. For US scans, clean-like targets are obtained via non-local low-rank (NLLR) denoising, removing the need for ground truth; for natural images, the originals serve as targets. Trained and validated on UDIAT~B, JNU-IFM, and XPIE Set-P, and evaluated additionally on a 700-image PSFHS test set, the method achieves the highest PSNR/SSIM across Gaussian and speckle noise levels, with margins that widen under stronger corruption. Relative to MSANN, Restormer, and DnCNN, it typically preserves an extra $\sim$1–4,dB PSNR and 0.05–0.15 SSIM in heavy Gaussian noise, and $\sim$2–5,dB PSNR and 0.05–0.20 SSIM under severe speckle. Controlled PSF studies show reduced FWHM and higher peak gradients, evidence of resolution recovery without edge erosion. Used as a plug-and-play preprocessor, it consistently boosts Dice for fetal head and pubic symphysis segmentation. Overall, the approach offers a practical, assumption-light path to robust US enhancement that generalizes across datasets, scanners, and degradation types.

💡 Research Summary

This paper tackles the long‑standing problem of ultrasound (US) image degradation, which stems from three intertwined physical factors: multiplicative speckle noise, blur introduced by the system point‑spread function (PSF), and scanner‑ or operator‑specific artifacts. While many deep‑learning‑based enhancement methods achieve impressive results, they typically rely on supervised training with clean ground‑truth images or on a known degradation model—assumptions that rarely hold in clinical practice.

The authors propose a blind, self‑supervised framework that simultaneously deconvolves and denoises B‑mode US frames. The core idea is to generate realistic training pairs directly from the available US data. For each frame, random rotations and crops produce patches (˜I). A physics‑guided degradation pipeline then synthesizes corrupted inputs (I_d) by (i) convolving ˜I with an isotropic Gaussian kernel that approximates the PSF (kernel size k ∈ {3,…,17}), and (ii) injecting two complementary noise processes: (a) spatial additive Gaussian noise (σ_g ∈ U(0.05,0.20)) to mimic thermal/receiver noise, and (b) complex Fourier‑domain perturbations (γ_f ∈ U(0,0.2)) that add zero‑mean complex Gaussian noise in the k‑space, reproducing phase and magnitude distortions characteristic of speckle. The order of blur and noise is randomly chosen (blur→noise with probability >0.55, otherwise noise→blur), creating a diverse corruption space and preventing the network from over‑fitting to a single artifact chronology.

Because true clean US images are unavailable, the authors generate “clean‑like” targets using a non‑local low‑rank (NLLR) denoiser applied to the original frames (I_t = D_NLLR(I)). NLLR exploits patch similarity across the whole image to produce a low‑rank approximation that retains structural content while suppressing speckle and blur. For natural‑image experiments (XPIE‑Set‑P), the original images serve as ground truth, allowing cross‑domain validation.

The enhancement network is a Swin‑Convolutional U‑Net (SC‑UNet). It blends local convolutional pathways with shifted‑window Swin‑Transformer self‑attention in hybrid blocks. Each block splits the channel dimension in half: one half passes through a lightweight 3×3‑ReLU‑3×3 residual convolution, the other half through a Swin‑Transformer windowed multi‑head self‑attention (window size 8, learnable relative positional bias). The two streams are concatenated, projected with a 1×1 convolution, and added back to the block input (residual connection). This design captures fine‑scale speckle statistics while also modeling long‑range contextual cues needed for deblurring. The encoder‑decoder follows a U‑Net topology with three down‑sampling stages (64→128→256→512 channels), a bottleneck, and three symmetric up‑sampling stages. Skip connections are additive rather than concatenative, reducing memory usage without sacrificing detail. A final residual connection with the stem features further preserves high‑frequency information.

Training minimizes an ℓ1 loss between the network output (\hat I = f_\theta(I_d)) and the target I_t. The ℓ1 loss is chosen for its robustness to outliers and its tendency to improve both PSNR and SSIM, especially under non‑Gaussian speckle. Training runs for 4000 epochs with a learning rate of 1e‑4, batch size 16, and on‑the‑fly generation of degraded inputs. Input patches are padded to the nearest multiple of 64 and cropped back after inference.

Evaluation is performed on three public US datasets—UDIA‑T (breast US), JNU‑IFM (intrapartum trans‑perineal US), and XPIE‑Set‑P (natural images used for cross‑domain exposure)—and on a dedicated 700‑image PSFHS test set that was never seen during training. Quantitative metrics include PSNR, SSIM, as well as resolution‑specific measures: full‑width at half‑maximum (FWHM), mean and maximum gradient, and contrast extracted from line profiles across anatomical edges.

Results show that the proposed method consistently outperforms state‑of‑the‑art baselines (MS‑ANN, Restormer, DnCNN). Under heavy Gaussian noise (σ_g = 0.20) the method gains 1–4 dB PSNR and 0.05–0.15 SSIM; under severe speckle (γ_f = 0.20) it gains 2–5 dB PSNR and 0.05–0.20 SSIM. Controlled PSF experiments demonstrate reduced FWHM and higher peak gradients, confirming genuine resolution recovery without edge smoothing.

Beyond image quality, the authors assess downstream impact by feeding the enhanced images to a UNet‑based segmentation model for fetal head and pubic symphysis delineation. Dice scores improve from 0.86 to 0.91 (fetal head) and from 0.78 to 0.84 (pubic symphysis), illustrating that the enhancement is not merely cosmetic but beneficial for clinical tasks.

The method generalizes across scanners (Siemens ACUSON, Y‑Probe) and anatomical regions, indicating robustness to device‑specific variations. Limitations include the focus on 2‑D B‑mode data; extending to 3‑D volumes or raw RF channel data would be a natural next step. Moreover, NLLR targets, while high‑quality, are not perfect ground truth, so residual artifacts may persist in very low‑signal regions.

In summary, this work presents a practical, assumption‑light pipeline that leverages physics‑guided synthetic degradations and self‑supervised learning to achieve state‑of‑the‑art ultrasound image enhancement. Its ability to improve both visual fidelity and downstream segmentation performance makes it a compelling candidate for integration into real‑world clinical imaging workflows.