ASCMamba: Adaptive Mamba Mechanisms

• ASCMamba is a context-dependent label for several adaptive Mamba-based designs applied in acoustics, medical imaging, and anomaly detection.
• It encapsulates variants that employ techniques like DenseEncoders, adaptive scan modules, and cross-attention to optimize task-specific performance.
• Empirical results demonstrate improved accuracy and segmentation metrics, highlighting benefits from metadata conditioning and adaptive scan-order learning.

Searching arXiv for the cited term variants and related papers.

ASCMamba is not a single uniformly defined architecture in recent arXiv literature. The label most directly names a multimodal acoustic scene classification system that combines a DenseEncoder, dual-path Mamba blocks, and conditional layer normalization for audio-text fusion (Sun et al., 21 Aug 2025). The same string is also used for the adaptive scanning module of ASM-UNet (Wang et al., 10 Aug 2025) and for the Mamba-enhanced cross-attention module inside MambaX-Net (Yahathugoda et al., 20 Oct 2025), while several adjacent papers identify “ASCMamba” as a misnomer or alternate rendering of differently named methods such as ASSM, ACMamba, and AS-Mamba (Zhang et al., 26 Mar 2025, Wang et al., 16 Apr 2025, Ning et al., 6 Feb 2026). This pattern suggests that “ASCMamba” functions less as a single canonical model name than as a context-dependent label for several adaptive or asymmetrical Mamba-based designs.

1. Terminology and disambiguation

The literature uses the string “ASCMamba” in at least three distinct ways: as an official model name, as the name of a module inside a larger architecture, and as a mistaken or alternate rendering of other model names. The following usages are explicitly documented.

Usage in the literature	Official designation	Application domain
ASCMamba (Sun et al., 21 Aug 2025)	"ASCMamba: Multimodal Time-Frequency Mamba for Acoustic Scene Classification"	Multimodal acoustic scene classification
ASCMamba in ASM-UNet (Wang et al., 10 Aug 2025)	Adaptive Scan Mamba	Fine-grained 3D medical image segmentation
ASCMamba in MambaX-Net (Yahathugoda et al., 20 Oct 2025)	Mamba-enhanced Cross-Attention Module	Longitudinal prostate MRI segmentation
“ASCMamba” as misnomer (Zhang et al., 26 Mar 2025)	Adaptive State-Space Mamba (ASSM)	Real-time sensor anomaly detection
“ASCMamba” as alternate rendering (Wang et al., 16 Apr 2025)	ACMamba	Hyperspectral anomaly detection
“ASCMamba” in AS-Mamba details (Ning et al., 6 Feb 2026)	AS-Mamba	CT metal artifact reduction

This terminological instability matters because the architectures, objectives, and datasets differ substantially. In one setting the term refers to multimodal audio-text fusion; in another it refers to adaptive scan-order learning over 3D feature volumes; in another it denotes a cross-attentional longitudinal fusion block. A careful reading therefore requires identifying the paper-specific definition before interpreting equations, claims, or results.

2. ASCMamba for multimodal acoustic scene classification

In its most literal usage, ASCMamba is the system introduced for the APSIPA ASC 2025 Grand Challenge, where acoustic scene classification is augmented with textual metadata such as location and record time (Sun et al., 21 Aug 2025). The model processes 10 s audio clips resampled to $f_s = 44.1$ kHz, uses a 40 ms Hann window, a 20 ms hop, an FFT size of $K = 2{,}048$, and $M = 64$ mel bands, and forms an LMFB input of $500 \times 64 \times 1$. A DenseEncoder extracts hierarchical spectral features and outputs a tensor $X \in \mathbb{R}^{B \times C \times T \times F}$. The architecture then applies two Mamba-based sequence paths: a time path using sequences $X_t \in \mathbb{R}^{(B \cdot F)\times T \times C}$ and a frequency path using sequences $X_f \in \mathbb{R}^{(B \cdot T)\times F \times C}$.

The multimodal fusion mechanism is Conditional Layer Normalization:

$$\mathrm{CLN}(x, c) = \gamma(c) \cdot \mathrm{LN}(x) + \beta(c),$$

where $c \in \mathbb{R}^{D_{\mathrm{cond}}}$ is a conditional vector projected from text embeddings. This design conditions both Mamba paths on metadata without using cross-attention. The paper describes this as an “early/mid” modulation strategy. The classification head fuses the time-path and frequency-path outputs and produces 10 scene logits.

Training couples supervised fine-tuning with a two-step pseudo-labeling pipeline. The procedure is: pretrain on a merged TAU Urban Acoustic Scenes 2020 Mobile and CochlScene set; fine-tune on the labeled portion of the CAS 2023 development set; pseudo-label the top 90% most confident unlabeled CAS clips with ASCMamba itself; use an improved SE-Trans baseline on the remaining 10% low-confidence clips; keep only class-agreement cases between ASCMamba and SE-Trans; and fine-tune again on the union of real labeled data and reliable pseudo-labeled data. The secondary pseudo-labeler uses two SE modules with multi-scale pooling, a Transformer encoder, and a two-head classifier with a 10-way scene posterior and a 2-way indoor/outdoor prior.

The reported behavior is highly dependent on metadata reliability. On Valid-Easy, ASCMamba with location and record time achieves average ACC 97.71%, compared with 93.75% without LRT and 92.73% for the SE-Trans baseline. On Valid-Hard(5%), however, ASCMamba with LRT drops to 94.14%, whereas the audio-only variant reaches 96.41%. On the official blind test set, the submitted top system is ASCMamba without LRT, reaching 64.4% macro-ACC, exceeding the baseline SE-Trans result of 58.2% by 6.2 percentage points and ranking first. The paper therefore treats multimodal conditioning as advantageous when metadata align with training conditions, but brittle under metadata mismatch or unseen cities.

3. ASCMamba as Adaptive Scan Mamba in ASM-UNet

In ASM-UNet, ASCMamba denotes “Adaptive Scan Mamba,” a module that learns an image-specific scanning order for 3D medical image features rather than relying on fixed horizontal or vertical sweeps (Wang et al., 10 Aug 2025). The motivating setting is fine-grained segmentation, particularly of small biliary structures whose diameters can be under 10 mm and whose topology varies substantially across patients. The parent architecture follows nnU-Net design principles with six encoder and six decoder blocks, and inserts ASCMamba after the first encoder block and before the last decoder block, where spatial resolutions are high and flattened sequence lengths are long.

For a feature map $F$ of shape $K = 2{,}048$0, spatial dimensions are flattened to a sequence of length $K = 2{,}048$1, producing $K = 2{,}048$2. The module then generates two score components. The group-level score is shared across the dataset and is obtained from a learnable vector interpolated to length $K = 2{,}048$3:

$K = 2{,}048$4

The individual-level score is inferred from three fixed-scan branches:

$K = 2{,}048$5

These are combined into a final adaptive score

$K = 2{,}048$6

which determines a permutation

$K = 2{,}048$7

used to reorder the flattened feature sequence before stacked Mamba layers are applied. The score is also concatenated to the reordered features, so the score generators influence downstream channel responses directly. The paper explicitly notes that no differentiable relaxation such as Gumbel-Softmax or Sinkhorn is used; ordering is obtained by vanilla $K = 2{,}048$8.

Empirically, ASM-UNet reports strong coarse-grained and fine-grained segmentation results. On ACDC, the average Dice is 92.61, which is +0.55 over nnFormer at 92.06. On Synapse, the average Dice is 87.27, which is +0.70 over nnFormer at 86.57. On the newly proposed BTMS biliary tract fine-grained segmentation dataset, the model achieves Avg. Coarse 76.74, Avg. Fine (w/o Hard) 48.79, and Avg. Fine (w/ Hard) 42.57, each reported as best among the compared methods. The hardest classes remain CD and RHD, where all methods stay below 35 Dice, indicating that adaptive scan-order learning improves but does not eliminate the difficulty of extremely small and highly variable ducts.

The ablation study attributes the gain specifically to the interaction between group commonalities and individual variations. The baseline UNet reaches 40.23 Dice on hard classes and 46.68 on non-hard classes. Adding only Mamba yields 40.85 and 46.97; adding Mamba with individual score yields 41.34 and 47.46; adding Mamba with group score yields 41.87 and 48.21; the full model reaches 42.57 and 48.79. A branch-count study further shows that 3 branches provide the best hard-class score, while 4 branches slightly improve non-hard classes.

4. ASCMamba as Mamba-enhanced cross-attention in longitudinal MRI

In MambaX-Net, ASCMamba refers to a Mamba-enhanced Cross-Attention Module designed for longitudinal Active Surveillance prostate MRI segmentation (Yahathugoda et al., 20 Oct 2025). The problem setting uses three inputs at time $K = 2{,}048$9: current T2-weighted MRI $M = 64$0, previous T2-weighted MRI $M = 64$1, and previous segmentation mask $M = 64$2 with three channels corresponding to whole prostate, peripheral zone, and transition zone. The architecture avoids explicit image registration and instead aligns current and previous observations in latent space.

MambaX-Net extends a 3D nnU-Net encoder-decoder with dual encoders and a single decoder. The current encoder processes $M = 64$3 through six downsampling stages, while the previous encoder processes $M = 64$4 through the top three stages; weights are shared across encoders. A Shape Extractor Module, implemented as three sequential 3D convolution plus ReLU blocks, encodes $M = 64$5 into a latent anatomical representation $M = 64$6. ASCMamba is inserted at the three highest decoder resolutions. At each such resolution, 3D feature maps are patch-embedded into sequences $M = 64$7, LayerNorm is applied, SEM features are fused after patch embedding, and two independent Mamba blocks process the current and previous sequences.

Cross-attention then aligns the two streams:

$M = 64$8

$M = 64$9

The attended sequence is unpatched back to a spatial feature map and fused residually:

$500 \times 64 \times 1$0

The paper emphasizes that Mamba supplies long-range spatial modeling at $500 \times 64 \times 1$1 complexity, whereas Transformer self-attention would scale as $500 \times 64 \times 1$2 in the number of patches.

Training is semi-supervised. A teacher nnU-Net pretrained on PI-CAI initializes the backbone and generates pseudo-labels for the private AS cohort. The student MambaX-Net is then fine-tuned on AS_train for 50 epochs using SGD, a OneCycle LR scheduler, batch size 2, early stopping after 15 epochs without validation improvement, and augmentations including random intensity shifts, flips, zooms, and Gaussian noise. The loss combines Dice plus Cross-Entropy with Focal-Tversky:

$500 \times 64 \times 1$3

The reported longitudinal gains are strongest in low-data settings. At $500 \times 64 \times 1$4 longitudinal patients, MambaX-Net achieves DSC WP/PZ/TZ of 0.91/0.85/0.87 and HD95 WP/PZ/TZ of 8.7/3.6/8.8 mm. At $500 \times 64 \times 1$5, DSC becomes 0.93/0.85/0.90 with HD95 7.2/4.8/7.6 mm. Using higher-quality pseudo-labels from SwinUNETR_PICAI at $500 \times 64 \times 1$6, MambaX-Net reaches DSC WP/PZ/TZ = 0.94/0.86/0.91 and HD95 WP/PZ/TZ = 7.1/3.4/7.6 mm. The ASCMamba ablation on AS_val shows that removing both SEM and Mamba collapses DSC to 0.57/0.46/0.63, whereas the full SEM+Mamba configuration yields 0.93/0.85/0.92. The study also reports that pre-registration of images provides no measurable benefit.

5. Misnomers, aliases, and adjacent usages

Several papers explicitly connect “ASCMamba” to methods whose official names differ.

ASSM for streaming sensor anomaly detection: In "Adaptive State-Space Mamba for Real-Time Sensor Data Anomaly Detection" the official model name is ASSM, and the paper states that “ASCMamba” is a misnomer in the query context (Zhang et al., 26 Mar 2025). ASSM defines the hidden-state update

$500 \times 64 \times 1$7

with

$500 \times 64 \times 1$8

and anomaly score

$500 \times 64 \times 1$9

Per-step complexity is reported as $X \in \mathbb{R}^{B \times C \times T \times F}$0 with memory $X \in \mathbb{R}^{B \times C \times T \times F}$1. On SWaT, ASSM reaches F1 = 0.80, ROC-AUC = 0.88, Latency = 4.1, and Throughput = 2,800 samples/s, outperforming KF, LSTM, and Transformer baselines.

ACMamba for hyperspectral anomaly detection: "ACMamba: Fast Unsupervised Anomaly Detection via An Asymmetrical Consensus State Space Model" notes that the model may appear as “ASCMamba,” but the official name is ACMamba (Wang et al., 16 Apr 2025). The method replaces dense pixel-level training with region-level instances obtained via SLIC superpixels, models region sequences with bidirectional RSAL Mamba scans, and introduces a Consensus Learning Strategy with two encoders and one shared decoder. The paper reports the highest mean AUC across eight benchmarks, approximately 0.9789, and the fastest mean detection time, approximately 1.50 s. On HyperMap, ACMamba reaches AUC approximately 0.8820 with training time approximately 4.17 s and inference time approximately 0.63 s.

AS-Mamba for CT metal artifact reduction: "AS-Mamba: Asymmetric Self-Guided Mamba Decoupled Iterative Network for Metal Artifact Reduction" is another nearby case in which the details explicitly equate ASCMamba with AS-Mamba (Ning et al., 6 Feb 2026). The architecture combines a high-frequency reduction network built from MambaBlocks, a Dual Enhancement Network for Fourier-domain amplitude correction, and an Iterative MANet for decoupled artifact and image refinement. On the synthetic DeepLesion benchmark, AS-Mamba reports average PSNR/SSIM of 45.04/0.9918, the best among the listed methods. On the First Affiliated Hospital clinical dataset, it reports STD = 31.25 HU, CNR = 0.98, and Clinical Score = 4.00.

MAMCA in automatic modulation classification: The paper "MAMCA -- Optimal on Accuracy and Efficiency for Automatic Modulation Classification with Extended Signal Length" explicitly frames its Mamba-backed AMC system as an “ASCMamba” in the concluding mapping from search terminology to model design (Zhang et al., 2024). MAMCA couples a denoising-capable soft-threshold unit with a Selective SSM backbone and reports average accuracy 89.67 on TorchSig-QAM, with 1.61 train time per epoch and 0.246 inference time per test set at signal length $X \in \mathbb{R}^{B \times C \times T \times F}$2.

6. Shared design patterns and recurrent limitations

Across these papers, ASCMamba-like systems consistently use Mamba or selective state-space modeling as a linear-time substitute for full-sequence attention, but they rarely deploy it in isolation. The official ASCMamba for acoustic scene classification couples dual-path Mamba blocks with DenseEncoder features, CLN-based metadata conditioning, and pseudo-labeling; ASM-UNet couples Mamba with adaptive scan scores derived from group commonalities and individual variations; MambaX-Net combines Mamba with longitudinal cross-attention and a shape prior; ACMamba combines region-level aggregation, RSAL bidirectional scans, and consensus learning; and AS-Mamba combines Mamba with wavelet decomposition, Fourier amplitude correction, and iterative decoupled refinement (Sun et al., 21 Aug 2025, Wang et al., 10 Aug 2025, Yahathugoda et al., 20 Oct 2025, Wang et al., 16 Apr 2025, Ning et al., 6 Feb 2026). This suggests that the central role of “ASCMamba” in current usage is not merely to denote a backbone, but to denote a task-adapted state-space mechanism embedded within a broader inductive-bias pipeline.

The limitations are correspondingly domain-specific. In multimodal ASC, location and record-time conditioning improves Valid-Easy accuracy but becomes brittle under metadata mismatch. In biliary fine-grained segmentation, the hardest classes remain difficult even with adaptive scanning. In longitudinal MRI, generalization is constrained by single-institution data and by sensitivity to pseudo-label quality. In hyperspectral detection, region-level aggregation assumes homogeneous superpixels and can dilute rare subtle anomalies. In CT metal artifact reduction, the method relies on supervised paired data and may retain residual artifacts for extremely dense metals. Taken together, these results indicate that “ASCMamba” is best understood as a family of adaptive Mamba formulations whose effectiveness depends on how successfully domain priors are encoded around the selective state-space core.