---
title: 'SFD-Mamba2Net: Dual-Use Architectures'
url: https://www.emergentmind.com/topics/sfd-mamba2net
type: topic
---

# SFD-Mamba2Net: Dual-Use Architectures

Searching arXiv for the cited SFD-Mamba2Net papers to ground the article in current records.
SFD-Mamba2Net is a name used on arXiv for two distinct Mamba-family architectures introduced in different application domains. In one usage, it denotes a **Mamba-based multimodal network** for **rapid blast-induced structural damage assessment (SDA)** from optical remote sensing imagery, with the 2020 Beirut explosion as the motivating case [2604.11709]. In the other, it denotes an end-to-end framework for **coronary artery segmentation and stenosis detection in invasive coronary angiography (ICA)** that combines structural priors, dual-stream state-space modeling, and frequency-domain reconstruction [2509.08934]. The shared name can suggest a single model lineage, but the two systems differ in task definition, modalities, module design, loss functions, datasets, and evaluation protocols.

## 1. Name, scope, and disambiguation

The label **SFD-Mamba2Net** is not tied to one canonical architecture. The 2026 blast-damage paper presents a multi-task multimodal network built around **Mamba / visual state space** components, inspired by **VMamba** and **ChangeMamba**, for **pixel-wise structural damage assessment** over an urban area after an explosion [2604.11709]. The 2025 medical-imaging paper presents an **asymmetric encoder-decoder** for ICA that targets **coronary artery segmentation and stenosis detection**, with modules named **CASE**, **AA-DS Mamba2**, and **PHFP** [2509.08934].

This naming overlap matters because the acronym “SFD” is domain-specific in each paper. In the blast-damage setting, the architecture is defined primarily by multimodal fusion of **pre-event imagery**, **post-event imagery**, and **blast-loading maps**. In the ICA setting, the architecture is defined by **structure-guided** enhancement, **frequency-aware** decoding, and a **dual-stream Mamba2** bottleneck. A plausible implication is that citations or repository references to “SFD-Mamba2Net” require domain context to avoid conflating urban damage mapping with coronary image analysis.

## 2. Structural damage assessment formulation

In the blast-damage paper, the target task is **pixel-wise structural damage assessment** after a major explosion. The model must infer **building localization / segmentation** from pre-event imagery, **damage classification** from pre- and post-event imagery, and may exploit **blast-loading information** that varies spatially with distance and exposure [2604.11709].

The problem formulation is motivated by the operational need for **fast, large-area, building-level damage maps** after an explosion. Classical field inspection is described as too slow and dangerous, while remote sensing provides a scalable alternative. The paper identifies three limitations in existing SDA systems: **dependence on labeled data**, **poor transfer to small local events**, and **lack of physical blast context**. To address these, it adopts a **two-stage pipeline** consisting of **pre-training** on **xBD** and **fine-tuning** on a small Beirut target dataset, while injecting **simulated blast-load maps** as an additional modality.

The core architecture contains four components: an **image encoder**, a **building-segmentation decoder**, a **blast-loading encoder**, and a **damage-assessment decoder** using a **Residual Attention–based Spatiotemporal State Space (RA-STSS)** fusion block. The optical inputs are the pre-event image $\mathbf{I}_{pre}$ and post-event image $\mathbf{I}_{post}$, and the third modality is the blast-loading map $\mathbf{B}$.

At the encoder front end, each optical image is patchified and embedded as
$$
\mathbf{F}_0 = \mathrm{PatchPartition}(\mathbf{I}) \in \mathbb{R}^{\tfrac{H}{4}\times\tfrac{W}{4}\times C_1},
$$
followed by hierarchical feature extraction with **VSS blocks** and downsampling:
$$
\mathbf{F}_1 = \mathrm{VSS}(\mathbf{F}_0), \qquad \mathbf{F}_l = \mathrm{VSS}(\mathrm{Down}(\mathbf{F}_{l-1})) \quad l=2,3,4.
$$
The encoder is applied independently to pre- and post-event imagery, producing multi-scale feature sets $\{\mathbf{F}_l^{pre}\}_{l=1}^4$ and $\{\mathbf{F}_l^{post}\}_{l=1}^4$. The paper attributes to Mamba/VSS efficient long-range spatial modeling with linear-time sequence processing while remaining lightweight enough for rapid deployment.

The first decoder reconstructs a binary building mask from pre-event imagery alone:
$$
\mathbf{M} = \mathrm{Decoder}(\{\mathbf{F}_l^{pre}\}_{l=1}^4) \in \mathbb{R}^{H \times W \times 3}.
$$
This auxiliary task supplies mask supervision so that the network learns building-aware features before damage classification. The blast modality is encoded separately as a spatial map $\mathbf{B}\in\mathbb{R}^{H\times W\times 3}$, interpolated to each pyramid scale and projected with a $1\times1$ convolution, yielding $\{\mathbf{F}_l^{blast}\}_{l=1}^4$.

The damage decoder first fuses pre- and post-event features:
$$
\mathbf{U}_l = \mathrm{STSS}(\mathrm{cat}(\mathbf{F}_l^{pre}, \mathbf{F}_l^{post})) \quad l=1,\dots,4,
$$
then performs a residual update modulated by the blast feature at the same level:
$$
\mathbf{D}_l = \mathbf{U}_l \oplus \mathrm{Up}(\mathbf{U}_{l-1}) * (1 + \mathbf{F}_l^{blast}) \quad l=1,\dots,4,
$$
with $\mathbf{U}_0 = 0$. The paper identifies this as the key multimodal fusion mechanism: the blast map acts as a **residual attention gate**, amplifying or suppressing features according to expected blast exposure. The final damage map is obtained by
$$
\mathbf{Y} = \mathrm{Argmax}(\mathbf{D}_4) \in \mathbb{R}^{H \times W \times T}.
$$

## 3. Blast-loading representation, training, and empirical results

A defining contribution of the blast-damage SFD-Mamba2Net is the explicit use of **multiscale blast-loading information** generated by physics-based simulation rather than inferred from imagery alone [2604.11709]. The paper states that the blast load is simulated using **Viper::Blast**, a GPU-based CFD solver validated for urban blast events. The Beirut explosion is approximated as a **0.50 kt TNT charge** detonated at a **height of 10 m above ground**, with a **cylindrical charge shape** and length-to-diameter aspect ratio 1.

Training is formulated as multi-task optimization with two pixel-wise cross-entropy terms, $\mathcal{L}_{mask}$ for building segmentation and $\mathcal{L}_{damage}$ for damage classification:
$$
\mathcal{L} = \mathcal{L}_{mask} + \mathcal{L}_{damage}.
$$
The target dataset, **Blast-7**, is constructed by combining **BRIGHT** with blast-loading data from the 2020 Beirut explosion and replacing BRIGHT’s post-event images with high-resolution optical imagery from the **Maxar Open Data Program**. Blast-7 contains **50 images** of **512 × 512 pixels**, covers about **7 km²**, and is split into **training / validation / test** in a **3:1:1** ratio. The pretraining dataset, **xBD**, contains **11,034** image pairs at **1024 × 1024** resolution, **850,736 buildings**, **19 disaster events**, and **45,361.79 km²** of coverage. The implementation uses **PyTorch**, **AdamW**, a learning rate of **$10^{-4}$** for pretraining, and **$10^{-5}$** for fine-tuning.

Evaluation follows the ChangeMamba/xBD protocol with $F_1^{\text{loc}}$, $F_1^{\text{clf}}$, $F_1^{\text{overall}}$, and class-wise $F_1$ for **Intact**, **Damaged**, and **Destroyed**. The reported best results are:

| Metric | Value |
|---|---:|
| $F_1^{\text{loc}}$ | 88.98% |
| $F_1^{\text{clf}}$ | 88.30% |
| $F_1^{\text{overall}}$ | 88.50% |
| Intact | 93.54% |
| Damaged | 77.96% |
| Destroyed | 95.64% |

The comparison set includes CNN-based baselines (**DeepLabV3+**, **U-Net**, **SiamAttnUNet**, **SiamCRNN**), a Transformer-based baseline (**DamFormer**), and a Mamba-based baseline (**Mamba-BDA-Small**). The paper reports that, compared with **UNet**, $F_1^{\text{overall}}$ improves by **27.55%**; compared with **SiamCRNN**, by **12.59%**; over **DamFormer**, overall classification improves by **7.3%** and damaged-class detection by **14.8%**; and compared with **Mamba-BDA-Small**, the method improves overall performance and especially the difficult **“damaged”** class by about **19.2%**. Fine-tuning requires only about **13 minutes**.

The ablation study isolates **pretraining**, **fine-tuning**, **distance information**, and **blast-loading information**. **Pretrain only** yields $F_1^{\text{overall}} = 24.52\%$ and $F_1^{\text{clf}} = 1.39\%$, indicating that global pretraining alone is insufficient for the Beirut scenario. **Fine-tuning (FT)** gives $F_1^{\text{overall}} = 85.98\%$, **FT with distance information (w/D)** gives **86.20%**, and **FT with blast loading (w/B)** reaches **88.50%**. This directly supports the paper’s claim that physics-informed blast features materially improve damage classification, particularly for the **intact** and **damaged** classes.

## 4. Coronary artery segmentation architecture

In the ICA paper, SFD-Mamba2Net is formulated as an end-to-end framework for **coronary artery segmentation and stenosis detection** in **invasive coronary angiography** [2509.08934]. The motivation is clinical: coronary artery disease is a leading cause of mortality, ICA is the gold standard for diagnosing coronary anatomy and stenosis severity, and manual interpretation is slow, subjective, and variable. Automated analysis is difficult because ICA images exhibit **low contrast**, **high noise**, **projection overlap**, **motion artifacts**, and extremely fine, tortuous vessel structures, especially **micro-vessels smaller than two pixels**, bifurcations, and distal branches.

The architecture uses an **asymmetric encoder-decoder**. The encoder has five stages, the first four built from conventional convolutional blocks, and the shallowest stage augmented by the **CASE** module. Each encoder block consists of two CBL units, where a CBL is **3×3 convolution + batch normalization + LeakyReLU**, with dropout after BN. The stage feature is defined as
$$
E_i = \text{LeakyReLU}\big(\text{BN}(\text{Conv}_{3\times3}(\text{LeakyReLU}(\text{BN}(\text{Conv}_{3\times3}(X_i)))))\big),
$$
and downsampling uses
$$
X_{i+1} = \text{LeakyReLU}\big(\text{BN}(\text{Conv}_{3\times3,\ \text{stride}=2}(E_i))\big).
$$
At the bottleneck, the model inserts **AA-DS Mamba2**, and the decoder mirrors the encoder with four upsampling stages. Each decoder stage concatenates the skip feature from the encoder with the upsampled output from the previous decoder stage,
$$
Y_i = \text{Concat}(U_{i+1}, E_i),
$$
where upsampling is
$$
U_{i+1} = \text{Conv}_{1\times1}(\text{Interpolate}(D_{i+1})).
$$
At the start of each decoder stage, the **PHFP** module refines the concatenated feature by progressive wavelet decomposition and reconstruction.

The **CASE** module, or **Curvature-Aware Structural Enhancement**, provides an explicit geometric prior for vessels. It uses a **multi-scale Hessian-based vesselness filter** inspired by Frangi-style vessel enhancement, with Gaussian smoothing at $\sigma \in \{1,2,3\}$, followed by Hessian computation
$$
H(X,Y) = \begin{bmatrix} D_{xx} & D_{xy} \\ D_{yx} & D_{yy} \end{bmatrix},
$$
and eigenvalues
$$
\lambda_{1,2} = \frac{D_{xx}+D_{yy}}{2} \pm \frac{1}{2}\sqrt{(D_{xx}-D_{yy})^2+4D_{xy}^2}.
$$
The vesselness response is
$$
\text{Vesselness}(x)=
\begin{cases}
\exp\!\left(-\dfrac{R_b^2}{2\beta^2}\right) \left(1-\exp\!\left(-\dfrac{S^2}{2c^2}\right)\right), & \lambda_2<0, \\
0, & \text{otherwise},
\end{cases}
$$
with
$$
R_b = \frac{|\lambda_1|}{|\lambda_2|+\epsilon}, \qquad S = \sqrt{\lambda_1^2+\lambda_2^2}.
$$
The paper emphasizes that CASE has **no learnable parameters**, takes the maximum vesselness response across scales, concatenates structural maps with early learned features, and applies a lightweight channel attention mechanism.

The **PHFP** module, or **Progressive High-Frequency Perception**, is designed to recover fine boundaries and small branches lost during downsampling. Using the **Haar wavelet**, it decomposes each decoder feature into $LL$, $LH$, $HL$, and $HH$ sub-bands:
$$
\text{WT}\big(Y^{(j-1)}_{LL}\big) \rightarrow \left(Y^{(j)}_{LL},Y^{(j)}_{LH},Y^{(j)}_{HL},Y^{(j)}_{HH}\right), \qquad j=1,\dots,n.
$$
High-frequency components are refined by a **$5\times5$ depthwise convolution** $D(\cdot)$, and reconstruction proceeds by inverse wavelet transform:
$$
R^{(j)} = \text{IWT}\!\left( Y^{(j)}_{LL} + D(Y^{(j)}_{LH}) + D(Y^{(j)}_{HL}) + D(Y^{(j)}_{HH}) + R^{(j+1)} \right), \qquad j=1,\dots,n-1.
$$
The final PHFP output is
$$
Y_{\text{PHFP}} = D(Y_i) + R^{(1)}.
$$
The paper’s interpretation is that low-frequency components preserve global vessel topology, while high-frequency components carry edges and fine branches.

## 5. AA-DS Mamba2, optimization, and segmentation performance

The ICA model’s global-context mechanism is the **Axial-Alternating Dual-Stream Mamba2 (AA-DS Mamba2)** bottleneck, which adapts **Mamba2** to 2-D vascular anatomy [2509.08934]. For input
$$
X \in \mathbb{R}^{B\times C\times H\times W},
$$
one path reshapes along width,
$$
X_{w1} = \text{Reshape}(X,\text{axis}=W)\in\mathbb{R}^{(B\cdot H)\times C\times W},
$$
applies DS Mamba2, reshapes back, then unfolds along height and applies DS Mamba2 again. The second path reverses that order. Their outputs are concatenated,
$$
Y_{\text{cat}} = \text{Concat}(Y_1,Y_2,\text{dim}=1)\in\mathbb{R}^{B\times 2C\times H\times W},
$$
and fused by
$$
Y_{\text{out}} = \text{Conv}_{1\times1}(Y_{\text{cat}}).
$$
The core DS Mamba2 block is bidirectional and processes the original sequence and its flipped version in parallel. The Mamba2 pipeline is described as input projection, activation and splitting into $x$, $B$, $C$, and $\Delta t$, SSD state propagation, fusion and normalization, and output projection. The structured state-space component is based on
$$
\dot{h}(t) = Ah(t) + Bx(t),\qquad y(t)=Ch(t),
$$
with
$$
A = -\exp(A_{\log}), \qquad \Delta t = \text{softplus}(\Delta t_{\text{raw}} + \Delta t_{\text{bias}}),
$$
followed by chunking into intra-chunk and inter-chunk terms, yielding $Y_{\text{diag}}$ and $Y_{\text{off}}$ before reshape and output.

Training uses **mean squared error (MSE)**,
$$
L_{\text{MSE}} = \sum_{i=1}^{N}(y_i-\hat{y}_i)^2,
$$
which the paper justifies as providing smooth gradients and helping suppress noise while preserving topology. Optimization uses **Adam** with initial learning rate **$1\times10^{-4}$**, $\beta_1=0.9$, $\beta_2=0.999$, and weight decay **$1\times10^{-5}$**. A **StepLR** schedule keeps the learning rate at **$10^{-4}$** for the first 10 epochs, halves it every 10 epochs until epoch 30, and then fixes it at **$3.125\times10^{-5}$**. Training is performed in PyTorch on **two RTX 3090 GPUs**.

The dataset comprises images from **99 patients** at Jiangsu Province People’s Hospital between Feb. 26 and July 18, 2019, acquired with a **Siemens AXIOM-Artis** system at **15 fps**, with resolution **$512\times512$** and spacing from **0.258 to 0.390 mm**. It contains **187 left coronary artery (LCA)** images and **127 right coronary artery (RCA)** images, annotated frame-by-frame by experienced interventional cardiologists. Segmentation is evaluated by **Dice**, **Acc**, **F1**, **IoU**, **Sens**, **Spec**, **HD95**, and **ASSD**.

The comparison set includes **U-Net, U-Net++, Swin-Unet, DSCNet, VM-UNet, Perspective-UNet, and ConDSeg**. The reported results are:

| Metric | Value |
|---|---:|
| Dice | 88.10% |
| F1 | 89.22% |
| IoU | 80.81% |
| Accuracy | 98.80% |
| Specificity | 99.50% |
| HD95 | 7.57 pixels |
| ASSD | 1.25 pixels |

Relative to **VM-UNet**, Dice improves from **87.30%** to **88.10%** and IoU from **80.06%** to **80.81%**, while **VM-UNet remained better** on **HD95** (**5.64** vs **7.57 pixels**). Relative to **ConDSeg**, Dice improves from **87.02%** to **88.10%**, F1 from **88.61%** to **89.22%**, IoU from **79.80%** to **80.81%**, and Specificity from **99.48%** to **99.50%**. The ablation study reports that adding the structural module alone gives modest improvement, **AA-DS Mamba2** improves Dice and reduces ASSD more substantially, **PHFP** improves boundary localization and especially lowers HD95, and the full model yields the best combined results.

## 6. Stenosis detection, qualitative interpretation, and limitations

In the ICA paper, stenosis detection is not framed as a separate end-to-end classification network; instead, it is a downstream pipeline operating on the segmentation output [2509.08934]. The vessel mask is first **skeletonized** to obtain the centerline, then the **Euclidean distance transform** is used to estimate local diameters:
$$
d(p) = 2\times \text{EDT}(p).
$$
The skeleton is decomposed into segments based on endpoint/bifurcation graph structure, and for each segment a stenosis ratio is computed as
$$
b = \left(1 - \frac{d_{\min}}{d_{\text{ref}}}\right)\times 100\%.
$$
A segment is marked as stenotic if $b \ge 0.1$ (10%). Additional constraints exclude segments with maximum diameter below **1.8 mm** and segments shorter than **20 pixels**; if no suitable inflection point is found, a global maximum diameter is used as reference. Severity is graded as **minimal: 1%–24%**, **mild: 25%–49%**, **moderate: 50%–69%**, and **severe: 70%–100%**.

The evaluation metrics are **TPR**, **PPV**, **ARMSE**, and **RRMSE**:
$$
\text{TPR} = \frac{TP}{TP+FN}, \qquad
\text{PPV} = \frac{TP}{TP+FP},
$$
$$
\text{ARMSE} = \sqrt{\frac{1}{N}\sum_{n=1}^{N}(b_e-b_g)^2}, \qquad
\text{RRMSE} = \sqrt{\frac{1}{N}\sum_{n=1}^{N}\left(\frac{b_e-b_g}{b_g}\right)^2}.
$$
Across **132 test images**, the paper reports **TPR = 0.60**, **PPV = 0.64**, **overall detection ratio = 0.4474**, **ARMSE = 0.14**, and **RRMSE = 0.30**. For **severe stenosis**, it reports **TPR = 0.34** and **PPV = 0.55**.

Both papers include qualitative evidence that is tightly coupled to their module design. In the blast-damage study, the salient qualitative claim is that adding **blast-loading information** provides physically grounded context beyond optical change alone, and the ablation results support that interpretation. In the ICA study, **Grad-CAM** visualizations are used to show stronger and more continuous activation on vessel structures for **CASE**, broader and more continuous attention over long vessel paths and bifurcations for **AA-DS Mamba2**, and sharper activations along branches and trunks for **PHFP**. This suggests that the two SFD-Mamba2Net variants, though unrelated in application, each use auxiliary structure to compensate for weaknesses of image-only feature extraction.

The limitations are likewise domain-specific. The blast-damage paper notes room for improvement through **fine-tuning large vision/language models** and richer multimodal SDA pipelines, while implicitly depending on a sufficiently good pretrained foundation model, simulated blast-load estimation, and some local fine-tuning samples. The ICA paper states explicitly that the dataset is relatively limited in scale and diversity, that the stenosis pipeline is based on **static single-view ICA** rather than multi-view or dynamic sequences, and that the 2-D projection of a 3-D coronary tree can reduce quantitative accuracy. Future work there is directed toward dataset expansion, multi-view and time-resolved ICA, and 3-D reconstruction approaches.

The combined record therefore presents SFD-Mamba2Net not as a single architecture but as a recurring name for two technically distinct systems: one for **physics-informed rapid structural damage assessment** after urban explosions [2604.11709], and one for **structure-guided, frequency-enhanced coronary analysis** in ICA [2509.08934].

Source: https://www.emergentmind.com/topics/sfd-mamba2net