TinyUSFM-uLPIPS Ultrasound Metric

• The paper introduces TinyUSFM-uLPIPS as a full-reference perceptual metric that compares multi-layer token relations and channel co-activation statistics to capture clinically relevant changes.
• It employs a compact TinyUSFM model to compute local self-similarity matrices and global Gram matrices, providing a modality-specific evaluation beyond traditional metrics like PSNR and SSIM.
• The approach demonstrates superior task alignment and cross-organ consistency, yielding more stable rankings in segmentation and reconstruction assessments compared to natural-image based metrics.

Searching arXiv for the specified paper and closely related context. TinyUSFM-uLPIPS is a full-reference perceptual distance for ultrasound image assessment introduced in “Defining Robust Ultrasound Quality Metrics via an Ultrasound Foundation Model” (Huang et al., 21 Apr 2026). It is built on TinyUSFM, a compact ultrasound foundation model, and compares a clean reference image $x$ with a distorted or reconstructed image $y$ using multi-layer token relations and channel co-activation statistics rather than pixel-wise fidelity alone. Within the same framework, it serves as the full-reference counterpart to TinyUSFM-NRQ, which is designed for no-reference quality scoring when no clean baseline is available (Huang et al., 21 Apr 2026).

1. Definition and intended role

TinyUSFM-uLPIPS is defined as an ultrasound-native, full-reference perceptual metric. In the reported experiments, the reference $x$ is the original high-quality ultrasound image, and the comparison image $y$ is a distortion or reconstruction produced at matched PSNR levels. The metric is therefore intended to quantify how much clinically relevant information has been altered or lost relative to the original image (Huang et al., 21 Apr 2026).

Its central design premise is that ultrasound image quality cannot be characterized adequately by generic natural-image perceptual features or by fidelity metrics such as PSNR and SSIM alone. The reported framework explicitly targets modality-specific structure, including speckle patterns, anatomical boundaries, shadows, attenuation effects, and scan-conversion-related appearance. TinyUSFM-uLPIPS addresses this by operating in an ultrasound-trained feature space and by measuring changes in local token relations and global feature-channel statistics rather than direct raw feature differences (Huang et al., 21 Apr 2026).

The paper attributes four advantages to the overall TinyUSFM-based evaluation framework: task-linked quality, cross-organ comparability, PSNR-consistent sensitivity for the no-reference score, and clinical utility. Within that framework, TinyUSFM-uLPIPS is the full-reference component responsible for task-linked quality and cross-organ comparability, and it is also used in conjunction with TinyUSFM-NRQ in super-resolution evaluation (Huang et al., 21 Apr 2026).

2. Mathematical construction

For an input ultrasound image $x$, TinyUSFM is treated as a vision-transformer-style encoder that outputs tokenized feature maps at multiple transformer layers: $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$ where $T_\ell$ is the number of spatial patch tokens, $C_\ell$ is the channel dimension, and the leading token is a global $[\mathrm{CLS}]$ token. Removing the $[\mathrm{CLS}]$ token gives

$y$0

TinyUSFM-uLPIPS uses the patch-token matrices $y$1 as its core representation. A fixed subset of layers is used in all experiments: $y$2 For each $y$3, the method extracts $y$4 and $y$5, then performs row-wise $y$6 normalization to obtain

$y$7

The first component is a local token-relation term. For a token index $y$8, a local neighborhood $y$9 of radius $x$0 is defined, with

$x$1

in all experiments. The neighborhood token matrix is

$x$2

formed by stacking the normalized token vectors in $x$3. The local self-similarity matrix is then

$x$4

Because the rows of $x$5 are $x$6-normalized, this matrix encodes cosine-like similarities among neighboring tokens.

To obtain probability-like neighborhood relations, the method applies a row-wise Softmax with temperature

$x$7

The layerwise local structural distance is

$x$8

The second component is a Gram-matrix term over channels: $x$9 with corresponding distance

$y$0

The per-layer distance is the sum

$y$1

and the final metric is the layer average

$y$2

This yields a scalar full-reference distance in which larger values indicate greater perceptual discrepancy (Huang et al., 21 Apr 2026).

3. Relation to LPIPS and ultrasound-specific rationale

Classical LPIPS, such as VGG-LPIPS, typically extracts deep features from a natural-image backbone such as VGG-16, normalizes channels, computes per-spatial-location squared feature differences, and aggregates over spatial locations and layers. TinyUSFM-uLPIPS differs in two essential ways: it uses TinyUSFM instead of VGG, and it replaces direct feature comparison with local self-similarity matrices and global Gram statistics (Huang et al., 21 Apr 2026).

The paper argues that PSNR and SSIM are limited because they prioritize pixel-level fidelity and global structural similarity, overreact to global gain or intensity shifts that do not harm diagnostic interpretability, and can fail to detect small but critical localized clutter or shadowing. It also argues that VGG-LPIPS and related natural-image metrics are affected by domain mismatch: ultrasound speckle and textured patterns may be misinterpreted as noise, whereas structured occlusions or shadows may not be penalized appropriately. In addition, natural-image metrics and some ultrasound-specific image-quality methods may not scale well across organs or devices (Huang et al., 21 Apr 2026).

TinyUSFM-uLPIPS is designed to address those limitations through an ultrasound-native feature space and a relational formulation. The local self-similarity term emphasizes how nearby patches relate to each other, so structured artifacts such as clutter haze, ROI shadowing, or missing scanlines strongly perturb these relations even when global intensity changes are small. The Gram term captures channel co-activation patterns across the entire image and is described as relating to global acoustic texture statistics, including speckle patterns and overall tissue texture. The paper interprets this design as making the metric more robust to benign variations such as gain or depth-dependent amplification that alter absolute intensities while preserving clinically relevant structure (Huang et al., 21 Apr 2026).

4. Calibration to semantic task damage

The paper evaluates task alignment by measuring how well metric values track segmentation Dice-score degradation under controlled distortions. The datasets are DDTI thyroid nodule segmentation and Open Kidney Ultrasound (KidneyUS) capsule segmentation. Starting from clean images $y$3, the study applies eight distortion families to produce $y$4 at matched PSNR levels, including noise or texture corruption, blur or resolution loss, ultrasound-specific artifacts such as shadowing, specular clipping, and missing scanlines, and geometric deformation. A fixed segmentation model is then applied, and “task damage” is defined as the drop in Dice relative to the clean image (Huang et al., 21 Apr 2026).

Metric ranking is compared with Dice-drop ranking using Spearman’s rank correlation $y$5 and Kendall’s rank correlation $y$6. According to Figure 1, TinyUSFM-uLPIPS shows higher and more stable Kendall’s $y$7 than VGG-LPIPS for both thyroid and kidney segmentation anchors, particularly at PSNR $y$8 dB. At higher PSNR, such as $y$9 dB, where Dice drops become small and rankings become harder, TinyUSFM-uLPIPS remains more stable than VGG-LPIPS (Huang et al., 21 Apr 2026).

A specific qualitative result is emphasized for thyroid images at PSNR $x$0 dB. VGG-LPIPS ranks Additive Gaussian noise as more severe than ClutterHaze or ROI Shadow Occlusion, whereas ClutterHaze and ROI Shadow Occlusion cause larger Dice drops. TinyUSFM-uLPIPS instead ranks those ultrasound-specific structural artifacts as more severe than additive noise, matching the segmentation results. In the paper’s interpretation, this indicates calibration to ultrasound semantics rather than to photographic priors (Huang et al., 21 Apr 2026).

5. Cross-organ comparability and operational behavior

Cross-organ comparability is assessed by generating PSNR-matched degradations between $x$1 and $x$2 dB for each organ and distortion, ranking distortions by metric severity within each organ, and then computing Kendall’s $x$3 across organs. Figure 2(a) shows that TinyUSFM-uLPIPS yields higher Kendall’s $x$4 than VGG-LPIPS across PSNR $x$5–$x$6 dB. This means that the relative severity ordering of distortions is more stable across thyroid and kidney when TinyUSFM-uLPIPS is used (Huang et al., 21 Apr 2026).

The paper also examines score-scale stability by computing the interquartile range (IQR) of metric scores across organs for the same distortion family and PSNR. Figure 2(b) shows that TinyUSFM-uLPIPS exhibits smaller inter-organ IQR on multiple representative distortions, including speckle-related and resolution-related corruptions. The reported interpretation is that lower IQR corresponds to more tightly clustered numerical scores for a given objective quality level, making thresholds or targets easier to interpret in multi-organ pipelines (Huang et al., 21 Apr 2026).

Implementation is deterministic once a pretrained TinyUSFM model is available. For each image pair, the method performs one TinyUSFM forward pass per image, extracts layers $x$7, removes the $x$8 token, applies row-wise normalization, computes neighborhood relation matrices with $x$9 and $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$0, forms Gram matrices, and averages the resulting distances across layers. No further finetuning is needed to train the metric itself. The paper notes that this is heavier than PSNR but practical for offline evaluation (Huang et al., 21 Apr 2026).

6. Use in reconstruction assessment, optimization, and reported limitations

TinyUSFM-uLPIPS is presented as applicable to reconstruction benchmarking, including super-resolution, denoising and speckle reduction, reconstruction algorithms such as beamforming variants and adaptive filtering, and generative models such as ultrasound image synthesis. In those settings it can replace or complement PSNR, SSIM, and VGG-LPIPS when a clean reference is available (Huang et al., 21 Apr 2026).

The same TinyUSFM feature space is also used as a differentiable perceptual loss for super-resolution: $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$1 In the reported super-resolution experiments, training SwinIR with $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$2 yields $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$3 PSNR and $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$4 SSIM over $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$5, together with better texture and fewer hallucinated structures than the VGG-based perceptual loss. TinyUSFM-uLPIPS and TinyUSFM-NRQ both score TinyUSFM-perceptual reconstructions more favorably, and clinician preferences follow the same direction: TinyUSFM-perceptual versus $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$6 is preferred in $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$7 of cases, TinyUSFM-perceptual versus VGG-perceptual in $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$8 of cases, and $f_\ell(x)\in\mathbb{R}^{(1+T_\ell)\times C_\ell},$9 versus VGG-perceptual in $T_\ell$0 of cases (Huang et al., 21 Apr 2026).

The reported setup also suggests several limitations. As a full-reference metric, TinyUSFM-uLPIPS cannot be used when no clean baseline is available; that use case is delegated to TinyUSFM-NRQ. Its behavior depends on the coverage of the pretrained TinyUSFM feature space, so underrepresentation of a particular organ or rare artifact type could plausibly reduce alignment in that domain. Because the score averages over layers and tokens, extremely small lesions or subtle changes that do not strongly perturb relational or Gram statistics might be underweighted. In addition, direct clinician-ranking validation is reported primarily for TinyUSFM-NRQ and for super-resolution outputs, so the full-reference metric is validated mainly through task anchors and through agreement with no-reference scores and expert preferences in super-resolution (Huang et al., 21 Apr 2026).

Taken together, TinyUSFM-uLPIPS is a modality-aligned full-reference ultrasound metric that replaces direct deep-feature differencing with a combination of local token-relation statistics and global Gram statistics in an ultrasound-native foundation-model space. In the reported experiments, that formulation yields better calibration to Dice-score drops, more stable cross-organ severity rankings, and operational compatibility with optimization loops that produce reconstructions preferred by sonographers (Huang et al., 21 Apr 2026).