UniUltra: Universal Ultrasound Segmentation

• UniUltra is an interactive, parameter-efficient framework for ultrasound segmentation, enabling a single model to segment diverse anatomical structures across various ultrasound modalities.
• It employs a two-stage pipeline with a context-edge hybrid adapter for fine-tuning SAM2 and deep-supervised knowledge distillation to compress the encoder for efficient deployment.
• Experimental results highlight competitive Dice scores and substantial parameter reduction, using only 8.91% of SAM2’s parameters during fine-tuning and achieving a 94.08% compression for deployment.

Searching arXiv for the exact topic name to anchor the article in the relevant paper. arXiv search query: "UniUltra" UniUltra is an interactive, parameter-efficient adaptation framework for the Segment Anything Model 2 (SAM2) designed for universal ultrasound segmentation, where “universal” denotes a single model that segments multiple anatomical structures and pathologies across heterogeneous ultrasound modalities and devices rather than a task-specific segmenter for one organ or acquisition protocol. The framework addresses the domain disparity between natural images and ultrasound by combining a context-edge hybrid adapter (CH-Adapter) for parameter-efficient fine-tuning with deep-supervised knowledge distillation (DSKD) for aggressive encoder compression. In reported experiments, UniUltra achieves competitive performance using only 8.91% of SAM2’s parameters during fine-tuning, and its final compressed form reduces the parameter count by 94.08% compared to the original SAM2, while preserving interactive prompting and strong cross-dataset generalization (Li et al., 19 Nov 2025).

1. Problem formulation and scope

UniUltra is motivated by a specific failure mode of foundation-model transfer: SAM2 demonstrates strong universal segmentation on natural images, but its performance on ultrasound degrades because ultrasound images exhibit heavy speckle noise, low contrast, artifacts and acoustic shadows, blurred and discontinuous boundaries, and marked heterogeneity across devices and clinical protocols. These characteristics differ from the RGB natural-image statistics assumed by SAM2’s pretraining, so direct deployment leads to under-detected boundaries and over-smoothed masks (Li et al., 19 Nov 2025).

In this setting, “universal ultrasound segmentation” means that a single promptable model should segment many distinct targets, including organs, lesions, fetal head, cervix, and vessels, across multiple modalities and scanners. The framework therefore targets not only segmentation accuracy but also promptability and deployment feasibility. The prompting regime is inherited from SAM2: bounding boxes, points, and combinations thereof are accepted, with bounding boxes emphasized during training because they simulate “touch” interactions on ultrasound devices. During training, bounding boxes are generated as the minimum enclosing rectangle of the mask with random perturbations of 0–20 pixels, a design intended to improve robustness to imprecise clinical input (Li et al., 19 Nov 2025).

A common misconception is to equate “universal” with prompt-free automation. The reported prompt ablations instead show that UniUltra remains strongest in an interactive regime. No-prompt performance is materially weaker than performance with prompts, whereas 1 point + 1 bounding box yields the best reported Dice and Hausdorff Distance (HD), and even a single bounding box already provides strong performance (Li et al., 19 Nov 2025). This suggests that UniUltra should be understood as a universal interactive ultrasound segmenter rather than a fully automatic universal segmenter.

2. Two-stage architecture and model variants

UniUltra is organized as a two-stage pipeline. The first stage produces UniUltra-L, a parameter-efficiently adapted SAM2-L model. The second stage produces UniUltra-mini, a compressed model in which the heavy image encoder is replaced by a super-light student encoder distilled from UniUltra-L (Li et al., 19 Nov 2025).

In the adaptation stage, UniUltra starts from SAM2-L with a Hiera-L image encoder, retains only the first 3 Hiera stages, freezes the Hiera blocks, inserts CH-Adapters between stages, and fine-tunes only the adapters and the decoder on ultrasound data. The prompt encoder and mask decoder are reused from SAM2, including the same loss composition of Dice + focal. In the compression stage, UniUltra treats the adapted large encoder as a teacher and distills its representations into a much smaller Hiera-style encoder with embedding dimension 128, while keeping the SAM2 prompt encoder and decoder unchanged (Li et al., 19 Nov 2025).

Variant	Core configuration	Reported role
UniUltra-L	SAM2-L with frozen Hiera blocks, CH-Adapters, decoder fine-tuning	Strong model for research or high-end clinical GPUs
UniUltra-mini	Distilled Hiera-style student encoder with embedding dimension 128 and 0.86M encoder parameters	Ultra-light deployment model for embedded or edge settings

The distinction between parameter-efficient fine-tuning and deployment efficiency is central. Parameter-efficient fine-tuning reduces the number of trainable parameters, but by itself does not reduce the inference-time footprint of the original SAM2 encoder. UniUltra therefore separates adaptation efficiency from deployment efficiency: CH-Adapter addresses the former, while DSKD addresses the latter. This separation is explicit in the reported results, where UniUltra-L is the highest-performing variant and UniUltra-mini is the deployment-oriented variant (Li et al., 19 Nov 2025).

3. Context-Edge Hybrid Adapter

The CH-Adapter is the principal adaptation module. It is inserted after the frozen Hiera processing at each of the three encoder stages and produces an adapted representation

$H_l = P^l + F_l,$

where $P^l$ is a context-aware residual term and $F_l$ is an edge-enhanced feature map (Li et al., 19 Nov 2025).

The context-aware component is a bottleneck MLP adapter: $$P^{l} = \mathrm{Linear}^{l}_{up}\left( \mathrm{GELU}\left( \mathrm{Linear}^{l}_{down}(F_l) \right) \right).$$ This branch learns a low-rank ultrasound-specific residual on top of frozen SAM2 features. The reported implementation reduces the embedding dimension within the adapter from 768 to 192 in UniUltra-L to avoid over-parameterization on relatively small medical datasets. Functionally, this branch is intended to preserve generic SAM2 semantics while injecting task-specific contextual cues such as typical lesion morphology and organ-level context (Li et al., 19 Nov 2025).

The edge-aware component is tailored to ultrasound boundary structure. After a patch-mixing reduction, the feature map is split channel-wise,

$F_{l} = \{F_l^i\}_{i=1}^{C},$

and each channel is filtered by four Sobel operators corresponding to horizontal, vertical, diagonal-right, and diagonal-left directions: $\gamma(S_k, F_l^i) = S_k * F_l^i.$ The multi-directional responses are aggregated,

$\sum_{k=1}^{n} \gamma(S_k, F_l^i), \quad n = 4,$

stacked across channels, and passed through a $1\times 1$ channel mixer $\mathcal{M}$: $$F_l \leftarrow \mathcal{M}\left(\left\{ \sum_{k=1}^{n} \gamma(S_k, F_l^i) \right\}_{i=1}^{C}\right).$$ This design reflects the observation that ultrasound edges are frequently weak, fragmented, and oblique rather than clean and axis-aligned (Li et al., 19 Nov 2025).

The reported parameter efficiency is explicit. Decoder-only fine-tuning uses 11.73M trainable parameters, a standard Adapter uses 20.66M, and CH-Adapter uses 20.80M, corresponding to approximately 8.91% of the baseline SAM2 parameters during fine-tuning. Despite similar scale to generic adapters, CH-Adapter yields better Dice and HD, which the authors attribute to the joint modeling of global context and directional boundary evidence (Li et al., 19 Nov 2025).

4. Deep-supervised knowledge distillation and compression

The second stage of UniUltra addresses a limitation common to parameter-efficient transfer methods: inference remains expensive if the original backbone is retained. UniUltra therefore introduces DSKD, which distills the adapted UniUltra-L encoder into a much smaller student encoder. The student follows a Hiera-style three-stage design but reduces the embedding dimension from 768 to 128, yielding an image encoder with only 0.86M parameters (Li et al., 19 Nov 2025).

The teacher exposes three types of intermediate signals at each stage $P^l$0: $P^l$1, the raw Hiera block outputs; $P^l$2, the CH-Adapter outputs; and $P^l$3, the integrated output formed by the element-wise addition of Hiera and CH-Adapter features. The student is trained to mimic all three streams. A small projection “neck” composed of a $P^l$4 convolution + $P^l$5 convolution aligns teacher and student feature shapes before matching (Li et al., 19 Nov 2025).

The DSKD objective is

$P^l$6

and the overall training objective is

$P^l$7

where the paper states the combined form but does not explicitly give a value for $P^l$8 (Li et al., 19 Nov 2025).

“Deep-supervised” refers to supervision at multiple hierarchy levels rather than only at the final encoder output. Reported ablations confirm that deeper distillation materially improves the compressed model. On TN3K, distillation at 1 layer yields Dice 92.34 and HD 54.66, while 3 layers yield Dice 93.99 and HD 35.47. This suggests that preserving the teacher’s multiscale representation, including both generic Hiera features and ultrasound-specific adapter features, is necessary for aggressive compression without substantial performance loss (Li et al., 19 Nov 2025).

5. Training protocol, datasets, and empirical performance

UniUltra-L is trained by initializing from SAM2-L weights, using the first three Hiera stages, freezing the Hiera blocks, and training the CH-Adapters and decoder. Inputs are resized to 1024 × 1024. The optimizer is Adam, with LR = 1e-4, exponential decay factor 0.98 per epoch, batch size 4, and 200 epochs. The implementation uses PyTorch 2.4.1, CUDA 12.4, and 4 × NVIDIA RTX A5000 (24GB) (Li et al., 19 Nov 2025).

The internal training corpus aggregates four public ultrasound segmentation datasets with an 8:1:1 train/validation/test split: BUSI (647 images), DDTI (637 images), TN3K (3493 images from 2421 patients), and UDIAT (163 images). External generalization is evaluated without fine-tuning on FUGC (500 images, 50 annotated) and JNU-IFM (6224 frames from 51 videos). Reported metrics are Dice, mIoU, HD, trainable parameters, image encoder parameters, FLOPs, FPS, and GPU memory consumption (Li et al., 19 Nov 2025).

On the four internal datasets, UniUltra-L reports:

• BUSI: Dice 93.81, mIoU 88.48, HD 42.67
• DDTI: Dice 94.20, mIoU 89.13, HD 43.93
• TN3K: Dice 94.50, mIoU 89.70, HD 34.27
• UDIAT: Dice 94.15, mIoU 89.06, HD 36.98

UniUltra-mini reports:

• BUSI: Dice 93.21, HD 49.13
• DDTI: Dice 94.04, HD 45.12
• TN3K: Dice 93.99, HD 35.47
• UDIAT: Dice 93.82, HD 40.41

The paper states that the improvement is statistically significant, with Dice p-values < 0.005 versus competitors and < 0.001 against MedSAM for UniUltra-mini (Li et al., 19 Nov 2025).

On external zero-shot validation, UniUltra-L reports:

• FUGC-AL: Dice 75.22, mIoU 62.86, HD 158.26
• FUGC-PL: Dice 82.07, mIoU 69.92, HD 138.49
• JNU-IFM-Head: Dice 96.37, mIoU 93.71, HD 19.44
• JNU-IFM-SP: Dice 90.83, mIoU 84.17, HD 13.83

UniUltra-mini reports:

• FUGC-AL: Dice 74.33, HD 170.45
• FUGC-PL: Dice 79.43, HD 140.27
• JNU-IFM-Head: Dice 94.95, HD 23.32
• JNU-IFM-SP: Dice 89.89, HD 15.72

These results are presented as evidence of cross-organ, cross-device, and cross-center generalization, especially relative to direct SAM or SAM2 transfer (Li et al., 19 Nov 2025).

The computational results clarify the system’s deployment profile. UniUltra-L reports 230.59M parameters, 847.47G FLOPs, 16.21 FPS, and Dice 90.14. UniUltra-mini reports 0.86M encoder parameters, 7.38G FLOPs, 45.09 FPS, and Dice 89.21. Relative to original SAM, UniUltra-mini reports parameter reduction of −99.86%, FLOP reduction of −99.73%, speed increase from 2.15 to 45.09 FPS, and Dice improvement of +13.56. Reported GPU memory usage is 9,881 MB for UniUltra-mini, compared with 33,821 MB for SAM2 and 77,083 MB for SAM (Li et al., 19 Nov 2025).

Interactive prompting remains a defining feature. The prompt ablation reports that no prompt yields Dice roughly 80–83% with poor HD, while 1 bounding box yields Dice roughly 93.8–94.5 and HD roughly 34–44. The best condition is 1 point + 1 bounding box; on TN3K it reports Dice 94.72 and HD 32.15 (Li et al., 19 Nov 2025).

6. Position within SAM adaptation research, limitations, and interpretation

UniUltra is positioned against three neighboring lines of work. First, it differs from generic SAM or SAM2 medical transfer methods because it explicitly targets ultrasound boundary degradation through a Sobel-based edge branch rather than relying only on low-rank or prompt-based adaptation. Second, it differs from parameter-efficient methods such as LoRA, Visual Prompt Tuning, and standard adapters because its CH-Adapter combines a bottleneck context branch with a dedicated edge-enhancement branch. Third, it differs from lightweight SAM variants such as MobileSAM, EfficientSAM, EdgeSAM, RepViT-SAM, and TinySAM because it compresses an already ultrasound-adapted teacher rather than distilling a generic foundation model directly (Li et al., 19 Nov 2025).

The framework’s clinical relevance follows from the combination of interactivity and resource efficiency. The reported 0.86M-parameter encoder, 45 FPS throughput at 1024×1024, and reduced memory footprint support deployment on ultrasound consoles, PACS workstations, or portable scanners. The use of bounding-box prompts is also presented as compatible with touchscreen and ROI-based workflows. This suggests that UniUltra is not merely a segmentation benchmark model but an attempt to reconcile foundation-model transfer with operational constraints in clinical ultrasound (Li et al., 19 Nov 2025).

Its limitations are explicit. All images are resized to 1024×1024, which may distort aspect ratios or small structures. The training corpus, while spanning four internal and two external datasets, does not cover all ultrasound modalities. Annotation quality constrains ceiling performance, especially on small datasets such as UDIAT and FUGC. Prompt-free segmentation is notably weaker than interactive segmentation. The framework is 2D-only, so volumetric ultrasound would require extending SAM2’s video or 3D capabilities. The paper also notes the risk of domain shifts outside the observed training distribution, including extreme noise, different probe frequencies, or atypical pathology (Li et al., 19 Nov 2025).

Taken together, UniUltra can be interpreted as a two-level answer to SAM2’s transfer problem in ultrasound. At the representation level, CH-Adapter modifies frozen SAM2 features by explicitly coupling global context with directional edge evidence. At the systems level, DSKD converts that adapted representation into an encoder small enough for practical deployment. The reported results indicate that the framework’s “universality” rests not on unrestricted modality coverage, but on the combination of multi-dataset training, promptable inference, and a compression strategy that preserves ultrasound-specific intermediate representations (Li et al., 19 Nov 2025).