Dual Hierarchical Contrast-Aware Mamba Module

• The paper introduces a dual hierarchical module that integrates global spatial and local temporal pathways to model multi-phase CT contrast variations.
• It employs innovative sequence designs, including phase-ordered scanning and similarity-guided refinement, to focus on diagnostically significant enhancement changes.
• Ablation studies confirm that coupling detail-preserving convolutions with multi-granularity fusion significantly boosts metrics like AUC and Recall.

Searching arXiv for the cited papers to ground the article. arxiv_search(query="(Wu et al., 18 Sep 2025)", max_results=5, sort_by="relevance") arxiv_search(query="(Gong et al., 16 Sep 2025)", max_results=5, sort_by="relevance") arxiv_search(query="(Ji, 5 Jun 2025)", max_results=5, sort_by="relevance") arxiv_search(query="Mamba 3D medical image segmentation SegMamba", max_results=10, sort_by="lastUpdatedDate") The Dual Hierarchical Contrast-Enhanced-Aware Mamba Module denotes a Mamba-based architectural pattern for medical imaging in which complementary sequence constructions and multi-scale feature pathways are used to encode both global context and localized contrast variation. In the literature provided, the term is defined explicitly as the core module of CECT-Mamba for multi-phase contrast-enhanced CT pancreatic tumor subtyping, where it combines a spatial hierarchy and a temporal hierarchy, both processed with Mamba, together with similarity-guided refinement and multi-scale aggregation (Gong et al., 16 Sep 2025). A closely related formulation appears in HybridMamba for 3D medical image segmentation, where the same conceptual structure is realized through slice-oriented versus local-adaptive scanning and spatial-versus-frequency gated fusion inside a hierarchical encoder-decoder (Wu et al., 18 Sep 2025).

1. Definition and conceptual scope

In CECT-Mamba, the module is described as a dual hierarchical contrast-enhanced-aware Mamba module incorporating two novel spatial and temporal sampling sequences to explore intra and inter-phase contrast variations of lesions (Gong et al., 16 Sep 2025). Its purpose is to model multi-phase contrast-enhancement patterns at two coupled levels. The first level is a global spatial hierarchy, which captures intra-phase spatial dependencies while respecting phase ordering; the second is a local temporal hierarchy, which tracks inter-phase enhancement changes at voxel or patch level across arterial, venous, and delayed phases. The module is therefore “dual hierarchical” because it couples two distinct but complementary representational hierarchies rather than relying on a single scan order or a single scale of context (Gong et al., 16 Sep 2025).

The phrase contrast-enhanced-aware has a specific operational meaning in this setting. It does not refer merely to using contrast-enhanced images as inputs; rather, it denotes architectural mechanisms that explicitly organize features according to phase-dependent enhancement variation. In CECT-Mamba this is achieved through phase-ordered sequence design and Similarity-guided Refinement (SimR), which emphasizes regions showing strong temporal change across phases (Gong et al., 16 Sep 2025). In HybridMamba, a related notion appears through FFT-based filtering and gated fusion of spatial and frequency-domain features, which explicitly extract both low-frequency shape cues and high-frequency edge or texture cues (Wu et al., 18 Sep 2025). This suggests that the term has both a narrow sense, tied to multi-phase CECT enhancement trajectories, and a broader architectural sense, tied to explicit modeling of contrast-sensitive feature variation.

A common misconception is to treat the module as a single standardized block with a fixed implementation. The available literature does not support that interpretation. The exact term is introduced in CECT-Mamba, whereas HybridMamba is described as something that can be understood as a concrete realization of the same idea for 3D segmentation, and DM-SegNet is presented as a close template rather than using the exact terminology (Gong et al., 16 Sep 2025, Wu et al., 18 Sep 2025, Ji, 5 Jun 2025). The concept is therefore best understood as a family of closely related Mamba modules organized around dual-path, hierarchical, and contrast-sensitive design principles.

2. Formal structure in CECT-Mamba

CECT-Mamba is designed for distinguishing pancreatic ductal adenocarcinoma from pancreatic neuroendocrine tumors using arterial, venous, and delayed CECT phases (Gong et al., 16 Sep 2025). After tumor localization and ROI cropping, phase-wise 3D feature volumes are extracted and passed into a 4-layer encoder whose core unit is the Dual Hierarchical Contrast-enhanced-aware Mamba module. The encoder output is then processed by a Multi-Granularity Fusion module and a classification head (Gong et al., 16 Sep 2025).

The DHCM operates on phase-wise features

$$F = \{ f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D} \}.$$

Each 3D feature volume is flattened into a 1D token sequence,

$$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$

after which two complementary scanning sequences are constructed (Gong et al., 16 Sep 2025).

The spatial embedding scanning sequence is formed by concatenating tokens within each phase and then concatenating the resulting phase sequences in phase order. This yields a phase-aware global volumetric sequence $\mathrm{x_s}$ that is processed as

$\mathrm{x_s' = Mamba\left(LN\left(\mathcal{M}(\mathrm{x_s})\right)\right), \quad \mathrm{x_s'' = MLP(\mathrm{x_s'})}.$

Here $\mathcal{M}$ denotes random masking with 50% tokens masked during training, and the resulting representation is intended to model global volumetric contrast variation under a phase-aware sequence structure (Gong et al., 16 Sep 2025).

The temporal embedding scanning sequence is constructed differently. For each spatial location $i$, the features from the three phases are linked in temporal order,

$[F_i^\mathrm{A}, F_i^\mathrm{V}, F_i^\mathrm{D}],$

and these local temporal chains are concatenated across positions to form a long 1D sequence $\mathrm{x_t}$. Temporal Mamba modeling is then written as

$\mathrm{x_t' = Mamba\left(LN\left(\mathrm{x_t}\right)\right), \quad \mathrm{x_t'' = MLP(\mathrm{x_t'})}.$

The outputs $\mathrm{x_s''}$ and $$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$0 are reshaped back to 3D and summed element-wise before downsampling to the next encoder stage (Gong et al., 16 Sep 2025).

This arrangement makes the module structurally distinct from naive multi-phase concatenation. The spatial branch asks, in effect, what the tumor looks like as a full 3D object across phases; the temporal branch asks how a fixed location changes through arterial, venous, and delayed enhancement. The two hierarchies therefore encode different invariants: one global and phase-conditioned, the other local and temporally dynamic (Gong et al., 16 Sep 2025).

3. Contrast-enhancement awareness and Similarity-guided Refinement

The most explicit contrast-enhanced-aware mechanism in CECT-Mamba is the Similarity-guided Refinement module. For each spatial location $$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$1, a similarity score is computed using cosine similarity between adjacent phases,

$$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$2

with

$$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$3

Low $$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$4 indicates strong variation across phases and is therefore taken as evidence of diagnostically important enhancement change (Gong et al., 16 Sep 2025).

Tokens are ranked by $$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$5 and split into high-similarity and low-similarity subsets. The low-similarity subset, corresponding to regions with stronger enhancement change, is passed through a Mamba-based refinement block and then restored to its original spatial positions before temporal sequence construction (Gong et al., 16 Sep 2025). In this design, SimR acts as a selective preprocessor that biases temporal modeling toward regions with clinically meaningful enhancement dynamics rather than expending capacity on static regions.

The ablation described for CECT-Mamba shows that adding SimR improves AUC (96.78%→97.22%) and Recall (90.60%→92.18%), which the authors interpret as confirmation that it emphasizes diagnostically useful temporal cues (Gong et al., 16 Sep 2025). This is important because it clarifies that “contrast-enhanced-aware” is not an ornamental label; it is instantiated as a measurable feature-selection and refinement mechanism.

A second misconception arises here. Contrast-awareness in DHCM is not implemented through a special loss tied directly to enhancement physics. The paper states that no special contrast-enhancement or similarity losses are defined; contrast-awareness is encoded via sequence design and SimR (Gong et al., 16 Sep 2025). The module is therefore contrast-aware by construction of feature order and refinement policy rather than by an explicit auxiliary optimization target.

4. Hierarchical encoding, spatial detail preservation, and multi-granularity fusion

DHCM is embedded in a broader hierarchical system that includes the Space Complementary Integrator and the Multi-Granularity Fusion module. These components are not secondary embellishments; they are part of how the dual hierarchies are stabilized and aggregated across scale (Gong et al., 16 Sep 2025).

SCI is introduced because Mamba’s flattening and sequence processing can lead to spatial detail loss. It processes a 3D feature map $$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$6 through two complementary convolutional paths, one using $$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$7 kernels for local spatial context and one using $$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$8 kernels for cross-channel context, before residual fusion: $$F = \{f^\mathrm{A}, f^\mathrm{V}, f^\mathrm{D}\} \;\longrightarrow\; S = \{ s^\mathrm{A}, s^\mathrm{V}, s^\mathrm{D} \},$$9 The reported ablation shows that adding SCI alone boosts AUC from 83.47% to 92.37% and Recall from 58.58% to 78.30%, indicating that preserving spatial detail before and around Mamba processing is a major contributor to performance (Gong et al., 16 Sep 2025).

MGF aggregates five hierarchical feature levels $\mathrm{x_s}$0 from shallow to deep encoder stages. Shallow features are aligned and fused through cross-scale multi-head self-attention, while deeper features undergo intra-scale self-attention before final cross-attention with the fused shallow representation (Gong et al., 16 Sep 2025). In functional terms, MGF transfers detail-rich shallow information into deeper, more abstract features and ensures that both micro-level patterns and macro-level shape or context contribute to subtyping. The full system achieves the best overall performance only after this cross-scale aggregation is added (Gong et al., 16 Sep 2025).

This hierarchical context matters for the meaning of the module itself. DHCM is not merely a pair of tokenization tricks attached to a classifier. It is a module that derives much of its effectiveness from being repeated within a multi-layer encoder and from being coupled to scale-aware feature aggregation. The “hierarchical” label therefore refers both to the internal dual hierarchy of spatial versus temporal scanning and to the external placement of the module within a multi-level encoder.

5. Segmentation-oriented realizations and related formulations

A closely related realization appears in HybridMamba, which is described as something that can be understood as a concrete realization of a Dual Hierarchical Contrast-Enhanced-Aware Mamba Module for 3D medical image segmentation (Wu et al., 18 Sep 2025). HybridMamba replaces the temporal hierarchy of multi-phase CECT with a dual scanning hierarchy inside a 3D encoder-decoder: SoMamba performs slice-oriented, global traversal, whereas LoMamba performs local-window, neighborhood-focused traversal. Their fusion is expressed in the Slice-Local Mamba block: $\mathrm{x_s}$1 SoMamba is defined as

$\mathrm{x_s}$2

and LoMamba as

$\mathrm{x_s}$3

where the three local sequences correspond to forward, reverse, and across-slice local traversal (Wu et al., 18 Sep 2025).

HybridMamba also introduces an FFT Gated Mechanism that makes the module explicitly spatial-frequency aware. Frequency-domain information is extracted by

$\mathrm{x_s}$4

with learnable thresholds for low- and high-frequency components, followed by gated fusion: $\mathrm{x_s}$5 The paper characterizes this as a dual-domain gating mechanism balancing spatial features and frequency-based features voxel-wise and channel-wise (Wu et al., 18 Sep 2025).

DM-SegNet offers a related but not terminologically identical formulation. It uses Mamba in both encoder and decoder, combining a quadri-directional spatial Mamba module, a gated spatial convolution layer, and a Mamba-driven decoding framework enabling bidirectional state synchronization across scales (Ji, 5 Jun 2025). In the provided analysis, this is presented as a concrete template for a dual hierarchical contrast-aware Mamba design, even though the exact phrase is not used in the paper itself (Ji, 5 Jun 2025).

Formulation	Dual hierarchy or duality	Contrast-aware mechanism
CECT-Mamba (Gong et al., 16 Sep 2025)	Spatial hierarchy + temporal hierarchy	SimR on low-similarity, high contrast-change regions
HybridMamba (Wu et al., 18 Sep 2025)	Slice/global scanning + local scanning; spatial + frequency domains	FFT filtering and gated spatial-frequency fusion
DM-SegNet (Ji, 5 Jun 2025)	Encoder Mamba + decoder Mamba; multi-scale fusion	Gated Spatial Convolution before QSMamba

This comparison shows that the module’s identity is architectural rather than strictly application-specific. In one case the dual hierarchy is spatial versus temporal across imaging phases; in another it is global slice traversal versus local window traversal; in a third it is encoder versus decoder Mamba coordinated across scales. What remains invariant is the pairing of complementary Mamba pathways with an explicitly detail-preserving or contrast-sensitive mechanism.

6. Empirical performance, interpretability, and limitations

In CECT-Mamba, the complete framework is evaluated on an in-house dataset of 270 patients (184 PDAC, 96 PNETs), each with arterial, venous, delayed phases → 810 3D volumes, using 5-fold cross-validation (Gong et al., 16 Sep 2025). The reported results are Accuracy 97.41% ± 2.77, AUC 98.60% ± 2.18, Recall 93.13% ± 6.88, Precision 98.06% ± 2.02, F1-score 95.48% ± 4.97, and Inference time 0.85 seconds per 10 cases. The model outperforms CNN, LSTM-based, Transformer, and Diff3Dformer baselines, and its AUC is reported as significantly higher with p < 0.05 in bootstrap testing (Gong et al., 16 Sep 2025).

The ablation trajectory in the same study is especially revealing because it isolates the contribution of each component. Starting from a basic 4-layer convolutional encoder, adding SCI, then spatial modeling, then temporal modeling, then SimR, and finally MGF leads to successive improvements, ending at ACC 97.41%, AUC 98.60%, Recall 93.13%, F1 95.48% (Gong et al., 16 Sep 2025). This pattern indicates that the module’s effectiveness depends not on a single ingredient but on the cumulative interaction of detail-preserving convolutions, dual Mamba sequence design, selective refinement, and hierarchical fusion.

HybridMamba provides analogous evidence in 3D segmentation. On BraTS2023, SegMamba achieves Avg Dice 91.32 and HD95 3.56, whereas HybridMamba reaches Avg Dice 91.92 and HD95 3.48; on the LC dataset, SegMamba reaches Dice 72.36, HD95 69.78, whereas HybridMamba reaches Dice 75.34 and HD95 44.52 (Wu et al., 18 Sep 2025). The ablation there shows that adding S-LMamba only improves Dice and HD95, adding FGM only improves them further, and combining both yields the best result, supporting the interpretation that dual scanning and spatial-frequency gating are complementary (Wu et al., 18 Sep 2025).

Interpretability evidence in CECT-Mamba is provided through Grad-CAM. In a challenging case where radiologists and nearly all comparison models misclassified a PNET as PDAC, CECT-Mamba correctly predicted PNET, and Grad-CAM indicated more precise focus on tumor peripheral regions and specific phase-wise enhancement regions (Gong et al., 16 Sep 2025). The paper argues that this focus mirrors radiological reasoning: comparing the same lesion location across phases while also integrating a global impression of the tumor.

The limitations are also clear. CECT-Mamba is evaluated on a single-center, in-house dataset, and the paper notes that external validation on multi-center cohorts is needed (Gong et al., 16 Sep 2025). It addresses only PDAC vs PNET subtyping, not broader pancreatic pathology, and it does not explicitly address robustness to motion artifacts or differing imaging protocols (Gong et al., 16 Sep 2025). More broadly, a plausible implication is that the term “Dual Hierarchical Contrast-Enhanced-Aware Mamba Module” currently names an emerging design pattern rather than a universally fixed module definition. Its exact implementation varies with task: multi-phase classification emphasizes temporal enhancement trajectories, whereas segmentation-oriented variants emphasize spatial scanning order, frequency-domain gating, or decoder-side state synchronization (Gong et al., 16 Sep 2025, Wu et al., 18 Sep 2025, Ji, 5 Jun 2025).