MAUP: Adaptive Uncertainty-Aware Prompting

• The paper introduces MAUP, a training-free strategy that leverages DINOv2 features and SAM prompting to perform cross-domain few-shot medical segmentation.
• The method constructs regional prototypes and uncertainty maps from support images via Voronoi partitioning and cosine similarity to generate effective prompts.
• Empirical results on Abd-MRI, Abd-CT, and Card-MRI demonstrate that MAUP improves Dice scores while reducing reliance on extensive annotated training data.

Searching arXiv for the MAUP paper and closely related uncertainty-aware prompting/domain adaptation papers. Multi-center Adaptive Uncertainty-aware Prompting (MAUP) is a training-free prompting strategy for Cross-domain Few-shot Medical Image Segmentation (CD-FSMIS) that adapts the Segment Anything Model (SAM), trained on natural images, to medical image segmentation through DINOv2-derived regional prototypes, multi-center point prompting, uncertainty-aware prompt selection, and adaptive prompt optimization (Zhu et al., 5 Aug 2025). It is proposed for the 1-way 1-shot setting, where a support image $I_s$ with binary mask $M_s$ is used to segment a query image $I_q$ without any additional training on source medical domains. The method is motivated by the observation that existing CD-FSMIS methods rely on heavy training over multiple source medical domains and carefully annotated datasets, whereas MAUP seeks to reduce deployment friction by replacing supervised adaptation with inference-time prompt engineering driven by robust features and region-aware statistics.

1. Problem setting and motivation

Cross-domain Few-shot Medical Image Segmentation aims to segment target classes in a target medical domain using only a few labeled examples, while the model is trained on source domains that can differ in modality or anatomy. The task is difficult because labels are scarce, expensive, and require clinical expertise; domain shift across modalities and institutions causes severe distribution gaps that degrade transferability; and existing CD-FSMIS methods rely on heavy training over multiple source medical domains and carefully annotated datasets, limiting universality and deployment ease (Zhu et al., 5 Aug 2025).

MAUP is designed as a training-free alternative. The rationale is that large visual foundation models trained on natural images, specifically SAM and DINOv2, provide powerful priors, but directly applying natural-image priors to medical images is non-trivial due to contrast ambiguities and anatomical complexity. MAUP therefore adapts SAM in a training-free manner by designing a prompting strategy that is driven by robust features and region-aware statistics rather than supervised optimization. This suggests that the principal innovation lies not in modifying the segmentation backbone, but in constructing prompts that translate cross-domain structural evidence into a form usable by SAM.

The method couples a frozen DINOv2-ViT-L/14 encoder with SAM-ViT-H. DINOv2 provides discriminative spatial features for support and query images, while SAM performs prompt-conditioned segmentation. In this arrangement, DINOv2 features are used to compute regional prototypes and similarity maps, and those maps are converted into multi-center, uncertainty-aware, adaptively optimized point prompts for SAM (Zhu et al., 5 Aug 2025).

2. End-to-end formulation

The MAUP pipeline begins with frozen feature extraction:

$F_s = E(I_s), \qquad F_q = E(I_q),$

where $E(\cdot)$ is a frozen DINOv2-ViT-L/14 encoder and $F_s, F_q \in \mathbb{R}^{C\times H\times W}$ (Zhu et al., 5 Aug 2025).

Support foreground information is then transformed into multiple regional prototypes by Regional Prototype Generation (RPG). The foreground defined by $M_s$ is partitioned using a Voronoi-based subdivision into $N_f$ regions, with $N_f = 30$. Let $\{V_n\}_{n=1}^{N_f}$ denote the regional masks. For each region, a prototype is computed by Masked Average Pooling:

$M_s$0

with $M_s$1 (Zhu et al., 5 Aug 2025).

Each support-region prototype induces a query similarity map through cosine similarity:

$M_s$2

This yields $M_s$3, each in $M_s$4 (Zhu et al., 5 Aug 2025).

MAUP also defines a periphery-based negative prototype. The support mask is dilated using a circular structuring element of radius $M_s$5:

$M_s$6

and the periphery mask is

$M_s$7

The periphery prototype is

$M_s$8

and its corresponding query similarity map is

$M_s$9

The numeric value of $I_q$0 is not specified (Zhu et al., 5 Aug 2025).

This construction produces three classes of evidence: multiple foreground-region similarities, their aggregate uncertainty structure, and a periphery-derived negative cue. A plausible implication is that MAUP treats the support mask not as a single prototype source, but as a structured object whose internal and peripheral organization can be projected into the query image.

3. Multi-center adaptive uncertainty-aware prompting

The core of MAUP is the transformation of similarity statistics into SAM point prompts. The mean similarity map is

$I_q$1

and the uncertainty map is the variance across regional similarities:

$I_q$2

High $I_q$3 indicates disagreement among regional prototypes and is associated with boundary ambiguity or regional inconsistency, particularly in low-contrast medical images (Zhu et al., 5 Aug 2025).

Positive prompts are generated through two paths. In the mean-similarity path, candidate points are defined by

$I_q$4

where $I_q$5 is the 95th percentile of $I_q$6. K-means clustering is then applied in pixel coordinate space to ensure spatial coverage and reduce redundancy. The K-means objective is

$I_q$7

with update rule

$I_q$8

Cluster centroids are used as positive prompt coordinates (Zhu et al., 5 Aug 2025).

The number of clusters is adaptive:

$I_q$9

$F_s = E(I_s), \qquad F_q = E(I_q),$0

Here, $F_s = E(I_s), \qquad F_q = E(I_q),$1 and $F_s = E(I_s), \qquad F_q = E(I_q),$2 are computed on the binarized high-similarity region, and $F_s = E(I_s), \qquad F_q = E(I_q),$3 is a scaling factor not numerically specified. This adaptively allocates more prompts for larger or more complex organs (Zhu et al., 5 Aug 2025).

In the uncertainty path, candidate points are

$F_s = E(I_s), \qquad F_q = E(I_q),$4

where $F_s = E(I_s), \qquad F_q = E(I_q),$5 is the 95th percentile of $F_s = E(I_s), \qquad F_q = E(I_q),$6. Two points are randomly selected from $F_s = E(I_s), \qquad F_q = E(I_q),$7 to focus on challenging regions such as boundaries and low-contrast areas. The final positive prompt set is

$F_s = E(I_s), \qquad F_q = E(I_q),$8

In the paper’s ablation terminology, uncertainty-map prompts (UMP) and mean-similarity prompts (MMP) are complementary components (Zhu et al., 5 Aug 2025).

Negative prompts are derived from the periphery similarity map:

$F_s = E(I_s), \qquad F_q = E(I_q),$9

where $E(\cdot)$0 is the 95th percentile of $E(\cdot)$1. These points are intended to lie near target periphery-like regions and help exclude surrounding tissues. The complete prompting scheme therefore combines multi-center coverage, uncertainty targeting, and periphery-aware exclusion (Zhu et al., 5 Aug 2025).

4. Integration with SAM and implementation characteristics

The prompts supplied to SAM are point coordinates with binary labels. Positive points in $E(\cdot)$2 are labeled as 1, and negative points in $E(\cdot)$3 are labeled as 0. MAUP uses SAM-ViT-H with point prompts only. The paper reports a single inference pass; no iterative re-prompting or post-processing such as CRF is reported, and the final mask comes directly from SAM’s inference conditioned on prompts (Zhu et al., 5 Aug 2025).

The end-to-end procedure is therefore: DINOv2 feature extraction for support and query, Voronoi partitioning of the support foreground into $E(\cdot)$4 regions, MAP-based prototype computation, cosine-similarity map generation, mean and variance estimation, thresholding and K-means clustering for positive prompts, random uncertainty-point sampling, periphery prototype construction for negative prompts, and a single SAM inference (Zhu et al., 5 Aug 2025).

The implementation details reported in the paper are limited but explicit in several respects. The framework is implemented in PyTorch v1.12 and evaluated on an NVIDIA RTX 3090 24GB GPU. The setting across all experiments is 1-way 1-shot. Thresholds $E(\cdot)$5, $E(\cdot)$6, and $E(\cdot)$7 are set to the 95th percentile of their respective maps. The adaptive prompt bounds are $E(\cdot)$8 and $E(\cdot)$9. Pre-processing is not explicitly detailed, and typical DINOv2 pipelines such as resizing and normalization are not specified in the paper. Exact feature-map resolution after DINOv2 extraction is also not specified (Zhu et al., 5 Aug 2025).

The qualitative complexity characterization is also reported. DINOv2 inference is described as $F_s, F_q \in \mathbb{R}^{C\times H\times W}$0 per encoder forward; similarity computation as $F_s, F_q \in \mathbb{R}^{C\times H\times W}$1; K-means on $F_s, F_q \in \mathbb{R}^{C\times H\times W}$2 candidate points for $F_s, F_q \in \mathbb{R}^{C\times H\times W}$3 clusters and $F_s, F_q \in \mathbb{R}^{C\times H\times W}$4 iterations as $F_s, F_q \in \mathbb{R}^{C\times H\times W}$5; morphological dilation as roughly $F_s, F_q \in \mathbb{R}^{C\times H\times W}$6 for a circular structuring element; and SAM inference as substantial but dominated by image resolution and model size. Exact runtime per case is not reported (Zhu et al., 5 Aug 2025).

5. Empirical performance and ablation evidence

MAUP is evaluated on three medical datasets: Abd-MRI (CHAOS), consisting of 20 abdominal MRI cases; Abd-CT (BTCV), consisting of 20 abdominal CT cases; and Card-MRI (MSCMR), consisting of 45 cardiac MRI cases. Evaluation uses Dice score (%) computed per 2D slice against 3D volumetric ground truth (Zhu et al., 5 Aug 2025).

The mean results reported for MAUP are summarized below.

Dataset	MAUP mean Dice	Improvement over second-best
Abd-MRI	67.09%	+1.30%
Abd-CT	67.46%	+2.71%
Card-MRI	73.13%	+1.55%

On Abd-MRI, the class-wise Dice scores for MAUP are 78.16 for Liver, 58.23 for LK, 72.34 for RK, and 59.65 for Spleen, with mean 67.09. The strongest baseline mean is FAMNet at 65.79, while SSL-ALP reaches 63.01, PATNet 52.97, RPT 46.91, IFA 40.61, and PANet 32.46 (Zhu et al., 5 Aug 2025).

On Abd-CT, MAUP reports 78.25 for Liver, 59.41 for LK, 71.80 for RK, and 60.38 for Spleen, with mean 67.46. The second-best mean is FAMNet at 64.75, followed by PATNet at 57.29, RPT at 48.28, SSL-ALP at 47.46, PANet at 31.94, and IFA at 30.79 (Zhu et al., 5 Aug 2025).

On Card-MRI, MAUP reports 88.36 for LV-BP, 52.74 for LV-MYO, and 78.29 for RV, with mean 73.13. The second-best mean is FAMNet at 71.58, followed by PATNet at 61.44, SSL-ALP at 57.47, RPT at 53.47, IFA at 37.50, and PANet at 36.66 (Zhu et al., 5 Aug 2025).

The ablation study on Abd-MRI mean Dice reports 65.08% for UMP only, 66.71% for UMP + MMP, and 67.09% for UMP + MMP + NP, where NP denotes negative periphery prompts. This shows that uncertainty-map prompts, mean-similarity prompts, and negative periphery prompts are complementary. A sensitivity study over the number of similarity maps indicates that performance varies with $F_s, F_q \in \mathbb{R}^{C\times H\times W}$7 and peaks at an optimal setting; the paper sets $F_s, F_q \in \mathbb{R}^{C\times H\times W}$8, although numeric peak values are not explicitly given in text (Zhu et al., 5 Aug 2025).

Visual comparisons across Abd-MRI, Abd-CT, and Card-MRI are reported to demonstrate improved boundary delineation and generalization across modalities and anatomies. Since the paper performs no additional training on source medical domains, these quantitative and qualitative results are presented as evidence that prompt engineering alone can outperform several training-based FSMIS and CD-FSMIS baselines in the reported setting (Zhu et al., 5 Aug 2025).

6. Related uncertainty-aware prompting, limitations, and extensions

MAUP belongs to a broader family of methods that use prompt design and uncertainty to adapt large pretrained models. A related medical-image segmentation framework is MedSAM-U, which introduces an uncertainty-guided auto-refining multi-prompt scheme for MedSAM. MedSAM-U differs structurally from MAUP because it trains a Multi-Prompt Adapter integrated with MedSAM, uses Uncertainty-Guided Multi-Prompt estimation with $F_s, F_q \in \mathbb{R}^{C\times H\times W}$9 prompt variants, and applies Uncertainty-Guided Prompt Adaptation that accepts a refinement only if uncertainty decreases (Zhou et al., 2024). MAUP, by contrast, is explicitly training-free, uses DINOv2-derived regional and periphery prototypes, and relies on single-pass prompting with point prompts only (Zhu et al., 5 Aug 2025).

A second related direction appears in universal multi-source domain adaptation rather than segmentation. APNE-CLIP formulates adaptive prompt learning with negative textual semantics and energy-based uncertainty modeling for Universal Multi-source Domain Adaptation. Its notion of “multi-source” domains can be interpreted as a multi-center setting, and its uncertainty-aware mechanism is based on energy separation between known and unknown samples rather than prompt-conditioned segmentation masks (Yang et al., 2024). This suggests that the phrase “multi-center adaptive uncertainty-aware prompting” has a broader conceptual range than the specific MAUP method for CD-FSMIS.

The limitations reported for MAUP are specific. Extremely low-contrast tissues or severe artifacts may still yield ambiguous masks, because some organs have indistinct boundaries and domain-specific textures that differ from natural images. The method depends on support mask quality, since RPG and the periphery prototype rely on an accurate support annotation. Several hyperparameters are heuristic, including the 95th-percentile thresholds, the dilation radius $M_s$0, and the scaling factor $M_s$1. The method does not explore iterative refinement, and evaluation is performed per 2D slice against 3D ground truth rather than through direct 3D volumetric prompting or segmentation (Zhu et al., 5 Aug 2025).

Potential extensions are stated as natural rather than definitive. A plausible implication is that iterative re-prompting with SAM feedback could improve boundary refinement, while direct 3D volumetric segmentation, multi-way tasks, and automatic parameter selection for $M_s$2, $M_s$3, and thresholding would address current scope limitations. The paper also highlights potential for real-world clinical applications, though explicit future work items are not listed (Zhu et al., 5 Aug 2025).

In summary, MAUP is a training-free CD-FSMIS approach that combines DINOv2-derived regional prototypes, multi-center similarity aggregation, uncertainty-driven prompt selection, periphery-aware negative prompts, and adaptive prompt quantity to condition SAM for medical image segmentation across domains (Zhu et al., 5 Aug 2025). Within the current literature, its distinguishing feature is the use of inference-time regional similarity statistics to replace source-domain training, thereby turning cross-domain adaptation into a prompt-construction problem rather than a parameter-update problem.