ASM-UNet: Adaptive Medical Image Segmentation

• The paper presents ASM-UNet, which replaces fixed scanning orders with data-driven adaptive scan scores to effectively capture both group-level anatomical norms and individual variations.
• The adaptive scanning mechanism leverages a U-Net backbone enhanced with an Adaptive Scanning Mamba block to improve feature extraction in fine-grained segmentation tasks.
• Empirical results on datasets like ACDC, Synapse, and BTMS show superior Dice scores, with ablation studies confirming the complementary benefits of combining group and individual scan scores.

Searching arXiv for the specified paper and closely related Mamba-based medical image segmentation context. ASM-UNet is a Mamba-based architecture for medical image fine-grained segmentation (FGS) that introduces adaptive scan scores to dynamically guide the scanning order by combining group-level commonalities and individual-level variations (Wang et al., 10 Aug 2025). It is motivated by clinical settings in which precise lesion resection depends on accurately identifying fine-grained anatomical structures, while frequent individual variations in small-scale anatomy make standard coarse-grained segmentation (CGS) approaches insufficient. Within this formulation, the central claim is that fixed manually-defined scanning orders, as used in prior Mamba-based models, limit adaptability in FGS, whereas a data-driven scan sequence can better align feature extraction with both shared anatomical structure and case-specific deviation (Wang et al., 10 Aug 2025).

1. Problem Setting and Motivation

The paper situates ASM-UNet in the gap between CGS and FGS. Many CGS methods have been successful in large-scale segmentation, such as organs, but they fall short in clinical scenarios requiring FGS because the target structures are small-scale and exhibit frequent individual variations (Wang et al., 10 Aug 2025). The summary further notes that traditional segmentation approaches, including both CNNs such as U-Net and Transformer-based models, may perform well for CGS but struggle with FGS for two stated reasons: frequent individual anatomical variations at small scales, and fixed scan orders in Mamba-based models that are not adaptive to anatomical diversity.

This framing makes the scanning order itself a modeling variable rather than a fixed implementation detail. In that sense, ASM-UNet treats sequence construction inside a Mamba or state-space model pipeline as part of the representation problem. A plausible implication is that the method is not only a backbone modification but also a redefinition of how spatial tokens are serialized before sequence modeling.

2. Network Organization

ASM-UNet is built on a U-Net backbone equipped with the Adaptive Scanning Mamba (ASM) block (Wang et al., 10 Aug 2025). The architecture is described as an encoder-decoder structure with 6 encoder and 6 decoder blocks. ASM blocks are placed after the first encoder and before the last decoder, specifically at positions where the feature maps are spatially large, that is, where the sequence length is long and the channel information is relatively limited, making global context extraction vital.

The network operates on a 3D medical image input of size $W \times H \times D$ and produces a segmentation map of size $W \times H \times D \times C$, where $C$ is the number of classes. Figure 1 in the paper is described as an overview of the proposed ASM-UNet architecture and illustrates the integration of ASM blocks into the U-Net pipeline.

The placement of ASM blocks is significant because it localizes the adaptive scanning mechanism to stages where long-range dependency modeling is most consequential. This suggests that the method is intended to intervene before heavy spatial compression removes the fine structural cues that FGS depends on.

3. Adaptive Scanning Mamba Mechanism

The Adaptive Scanning Mamba mechanism replaces the fixed scan order used in prior Mamba or state-space-model segmentation pipelines with an adaptive scan score computed per image and per spatial position (Wang et al., 10 Aug 2025). The summary gives the following process flow:

1. Spatial flattening: feature maps of shape $(W, H, D, L)$ are reshaped to $(W \times H \times D, L)$.
2. Adaptive scan score generation: a group scan score $g$ and an individual scan score $i$ are computed.
3. Score integration: the scores are summed and normalized.
4. Feature reordering: the concatenated scan-score-plus-feature representation is sorted according to the adaptive scan score.
5. Mamba feature extraction: the reordered sequence is processed via stacked Mamba layers.
6. Inverse reordering and aggregation: outputs are restored to the original spatial order and pixel-wise averaged.
7. Reshaping: the result is reshaped spatially and passed to the decoder path.

The group scan score is described as a learned, spatially-broadcastable embedding reflecting population-level anatomical patterns. The individual scan score is dynamically derived from input-specific features using three fixed scan orders and is described, in the mathematical summary, as being computed via self-attention over several canonical scan orderings. The adaptive scan score is presented in two closely related forms in the summary:

$$\text{adaptive\_scan\_score} = \text{group\_scan\_score} + \text{individual\_scan\_score}$$

and

$a_s = \sigma(g_s + i_s),$

where $\sigma$ is the sigmoid function.

The scan order is then determined by sorting $W \times H \times D \times C$0, and the feature $W \times H \times D \times C$1 is concatenated with $W \times H \times D \times C$2 before sequence modeling via Mamba. Figure 2 is described as showing adaptive score generation and the sorting mechanism.

4. Group Commonalities and Individual Variations

A defining feature of ASM-UNet is the explicit integration of group commonalities and individual variations (Wang et al., 10 Aug 2025). The group scan score captures general anatomical layout and is intended to model expectable, stable features. The individual scan score leverages patient-specific image information and allows adaptive scanning to suit unique anatomical peculiarities. Their sum constitutes joint guidance for the downstream scan sequence.

In the paper summary, this joint guidance is described as dynamically prioritizing scan paths that are both generally appropriate and individually responsive, thereby improving attention to structures that may be small, misshapen, or variable. The same summary draws a clinical parallel: expert radiologists use both general anatomical knowledge and case-specific observations to guide their reading strategy. Within the model, that analogy corresponds to the coexistence of population-level priors and instance-conditioned modulation.

A plausible implication is that ASM-UNet treats FGS as a problem in which the optimal ordering of evidence is anatomy-dependent rather than fixed across a dataset. This is distinct from merely increasing model capacity, because the mechanism changes the traversal strategy used to expose features to the state-space module.

5. Empirical Evaluation

The experiments are reported on two public datasets, ACDC and Synapse, and on a newly proposed challenging biliary tract FGS dataset, BTMS (Wang et al., 10 Aug 2025). The Dice Similarity Coefficient is identified as the primary metric for all experiments.

Dataset	Composition	Reported result
ACDC	Cardiac MRI; 3 classes; 2,978 slices (150 cases)	92.61% average Dice
Synapse	Abdominal CT multi-organ; 8 classes; 3,779 slices (30 cases)	87.27% average Dice
BTMS	8 categories; 100 cases (20,419 images)	76.74% coarse; 48.79% fine excluding hard; 42.57% fine including hard

For ACDC and Synapse, ASM-UNet achieves the top average Dice, specifically 92.61% on ACDC and 87.27% on Synapse. On ACDC it is reported as best in RV and LV and competitive in MYO; on Synapse it is reported as best or second-best in nearly all organs. For BTMS, the reported scores are 76.74% for Avg. Coarse, 48.79% for Avg. Fine excluding hard classes, and 42.57% for Avg. Fine including hard classes. The summary states that these results beat second-best methods by significant margins in all metrics. On individual categories such as GB, CBD, and CHD, ASM-UNet is reported to rank highest; for the most challenging classes, CD and RHD, all methods struggle, but ASM-UNet shows notable improvements.

The ablation study isolates the contribution of the scan-score components. Adding the individual scan score improves fine Dice from 40.85% to 41.34%; adding the group scan score improves it further to 41.87%; and using both individual and group scan scores yields the highest value at 42.57% with hard classes included. The summary interprets this as a complementary impact: each module is independently useful, but performance is best when they are combined. A separate parameter analysis reports that multiple adaptive scan branches, especially 3, improve performance, indicating that diverse scan order explorations aid representation.

6. Qualitative Behavior, Impact, and Limits

The qualitative results, described for Figures 7 through 9, report that ASM-UNet produces cleaner, more anatomically-plausible, and less fragmented segmentations on FGS, especially for small or ambiguous structures (Wang et al., 10 Aug 2025). In hard-class segmentation such as CD and RHD, performance remains limited for all models because of anatomical variability and scale, but ASM-UNet is reported to show superior continuity and fewer false positives.

The summary attributes several effects to adaptive scan scores: dynamic, data-driven scan orders per image and region; more contextually-aware feature extraction in challenging and variable anatomy; and improved model generalizability and robustness, supported by consistent gains across the public datasets and BTMS. In this framing, the adaptive guidance mechanism is not only a local architectural modification but the principal explanation for improved fine-scale partitioning, reduced misses, and reduced over- or under-segmentation in regions with ambiguous boundaries or unusual spatial configurations.

At the same time, the BTMS results on hard classes show that adaptive scanning does not eliminate the intrinsic difficulty of highly variable, small-scale anatomical targets. That limitation is explicit in the reported findings: even with ASM-UNet, CD and RHD remain challenging for all methods.

7. Position Within Fine-Grained Medical Segmentation

ASM-UNet is presented as an advance in Mamba-based medical image segmentation through adaptive, dual-level scan score guidance that leverages both group anatomical norms and individual image nuances (Wang et al., 10 Aug 2025). In the terminology of the paper, the Adaptive Scan Mamba mechanism is pivotal because it makes feature learning relevant to the specific case while retaining access to group-level anatomical regularities.

Within the scope of the reported evidence, the model is characterized as achieving superior performance in both CGS and FGS tasks on ACDC, Synapse, and BTMS. The BTMS dataset is introduced as a challenging biliary tract FGS benchmark with 8 categories and 100 cases comprising 20,419 images, which places particular emphasis on small-scale anatomical targets and individual variability. The accompanying resources are reported as publicly available at the project repository: https://github.com/YqunYang/ASM-UNet.

Taken together, the method’s distinctive contribution is the replacement of fixed manually-defined scanning orders with adaptive scan scores derived from both shared and individual anatomical structure. This suggests a broader view of fine-grained segmentation in which the order of sequence exposure is itself a learnable anatomical prior rather than a fixed preprocessing choice.