Mammo-Mamba: Hybrid Mammogram Model

• Mammo-Mamba is a hybrid model for multi-view mammogram classification that integrates selective state-space models, transformer self-attention, and SeqMoE gating.
• It employs a four-stage MambaVision backbone with dual-stream fusion to capture both local lesion details and global anatomical context.
• The architecture delivers improved performance (Accuracy 0.8792, AUC 0.9249) while achieving linear computational scaling versus traditional Vision Transformers.

Searching arXiv for the specified paper and closely related work on mammography and Vision Mamba architectures. Mammo-Mamba is a hybrid architecture for multi-view mammography that integrates Selective State-Space Models (SSMs), Transformer self-attention, and a Sequential Mixture-of-Experts (SeqMoE) gating mechanism within a four-stage MambaVision backbone. It is designed for settings in which mammographic screening uses cranio-caudal (CC) and medio-lateral oblique (MLO) views and may examine both whole images and lesion-cropped regions. The model replaces most self-attention layers with linear-time SSM layers, retains a smaller number of self-attention blocks to refine local/global interactions, and uses dual-stream fusion of cropped and whole mammograms in a unified, end-to-end learnable pipeline (Bayatmakou et al., 23 Jul 2025).

1. Clinical and modeling context

Breast cancer remains one of the leading causes of cancer-related mortality among women, and accurate as well as efficient interpretation of multi-view mammograms is essential for early detection. The design problem addressed by Mammo-Mamba is therefore not only classification accuracy but also computational scaling: state-of-the-art multi-view mammogram classification models are described as largely based on Transformer architectures, while the quadratic complexity of full self-attention with respect to patch count becomes problematic for high-resolution mammographic images (Bayatmakou et al., 23 Jul 2025).

The motivating requirements are stated explicitly. A suitable model should scale to thousands of patches with long-range dependency modeling beyond local convolution, avoid the quadratic patch–patch cost of vanilla Vision Transformers, adapt its computational effort dynamically based on image content, and fuse multiple views and magnifications in a unified pipeline. Mammo-Mamba is presented as meeting these goals by replacing most self-attention layers with SecMamba blocks, retaining a smaller number of self-attention layers, and applying SeqMoE gating so that each block can decide whether to apply a shallow or deeper transformation (Bayatmakou et al., 23 Jul 2025).

This design emphasis reflects a mammography-specific tension between fine-grained tumor morphology and broader breast anatomy. The paper’s dual-stream formulation, based on cropped and whole mammograms, is intended to preserve both local lesion evidence and global contextual structure rather than privileging only one scale of representation.

2. Backbone organization and multi-stage composition

Mammo-Mamba is nested within the four-stage MambaVision backbone. The architecture preserves the hierarchical staging of MambaVision but customizes the deeper stages with SecMamba and SeqMoE. The first two stages are convolutional residual blocks that are frozen to preserve general image priors, whereas the later stages contain the task-specialized sequence modeling and gating machinery (Bayatmakou et al., 23 Jul 2025).

The stage-wise organization reported for Mammo-Mamba is as follows:

Stage	Components	Function
1–2	convolutional residual blocks, frozen	preserve general image priors
3	ten experts in series: 5 SecMamba blocks interleaved with 5 self-attention blocks, each followed by SeqMoE	progressive emphasis in deeper features
4	spatial downsampling via strided convolution, then five experts: 3 SecMamba and 2 self-attention, each also gated by SeqMoE	deeper specialization for mammographic patterns

The paper describes this as “progressive emphasis.” Low-level texture and edge features remain general, while deeper layers specialize in mammographic patterns and adjust emphasis on a per-patch basis as data flows through successive expert transformations. This suggests that the architecture is intended to separate generic visual priors from content-adaptive mammography-specific refinement, rather than re-learning all stages end to end from scratch.

A central architectural consequence is that self-attention is not eliminated. Instead, it is sparsified across the depth of the network: Stage 3 contains five self-attention blocks and Stage 4 contains two. The long-range processing burden is shifted primarily to SSM layers, while attention remains available for additional local–global interaction refinement.

3. SecMamba and Sequential Mixture-of-Experts

At the core of Mammo-Mamba is the SecMamba block, a modified MambaVision block that combines a discrete-time linear state-space model with selective gating and residual projection. For the $k$-th patch in image $i$, the SSM is written as

$h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$

$y_{i,k} = C\,h_{i,k},$

where $h_{i,k}\in \mathbb{R}^{d_{HS}}$, and $A\in\mathbb{R}^{d_{HS}\times d_{HS}}$, $B\in\mathbb{R}^{d_{HS}\times d_{IE}}$, and $C\in\mathbb{R}^{d_{IE}\times d_{HS}}$ are learned. Rather than iterating these equations explicitly across time, the sequence of $N_P$ patches is processed through a single 1D convolution,

$$Y_i = K^T * X_i,\qquad K = [C\,B,\; C\,A\,B,\; \dots,\; C\,A^{N_P-1} B],$$

which yields an $i$0 pass that is linear in patch count rather than the $i$1 cost of full self-attention (Bayatmakou et al., 23 Jul 2025).

Selective gating is then applied:

$i$2

followed by projection and residual normalization,

$i$3

In practice, the paper states that $i$4–$i$5, while $i$6 is chosen to balance expressivity with efficiency.

SeqMoE is the second defining mechanism. Unlike classical MoE, which deploys parallel experts and gates across them, SeqMoE arranges each SecMamba or self-attention block in sequence and uses a gate to determine whether to use the block output directly or blend it with the incoming representation. The gate for layer $i$7 is

$i$8

and the resulting mixture is

$i$9

The paper characterizes this as allowing each patch to choose “shallow” versus “deeper” processing and as enabling content-adaptive feature refinement, with high-uncertainty regions routed through extra transformation while easier areas pass through unchanged (Bayatmakou et al., 23 Jul 2025).

4. Integration of attention, dual-stream fusion, and computational profile

Within Stages 3 and 4, SecMamba blocks alternate or interleave with standard multi-head self-attention layers. For a patch sequence $h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$0, the attention computation is given as

$h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$1

$h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$2

$h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$3

Because the number of attention blocks is small relative to the SSM blocks, the quadratic term is retained only with a small coefficient. The reported computational comparison is explicit: a standard Vision Transformer requires $h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$4 time and $h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$5 memory, whereas Mammo-Mamba uses $h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$6 time and $h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$7 memory for SecMamba SSMs, with SeqMoE gating treated as negligible at $h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$8 per layer (Bayatmakou et al., 23 Jul 2025).

The paper further states that empirical GPU memory footprints remain close to linear scaling with $h_{i,k} = A\,h_{i,k-1} + B\,x_{i,k-1},$9, and that the dominating cost shifts to SSM convolutions plus only a handful of attention modules. This is contrasted directly with a full Transformer on $y_{i,k} = C\,h_{i,k},$0 patches.

Dual-stream fusion is another structural element. Mammographic screening typically uses CC and MLO views and may analyze both whole images and lesion-cropped regions. Mammo-Mamba fuses cropped and whole mammograms so that local and global cues jointly guide classification. In the reported ablations, this dual-stream design materially changes performance relative to whole-only or crop-only baselines, which is consistent with the architectural motivation (Bayatmakou et al., 23 Jul 2025).

5. Training protocol and reported empirical performance

The reported evaluation is conducted on the CBIS-DDSM benchmark dataset, with 653 examinations using CC/MLO and whole plus crop inputs, split into 504 training and 149 test cases. Preprocessing consists of resizing to $y_{i,k} = C\,h_{i,k},$1, random color jitter, and horizontal flip for view invariance. Training uses AdamW with learning rate in $y_{i,k} = C\,h_{i,k},$2, weight decay in $y_{i,k} = C\,h_{i,k},$3–$y_{i,k} = C\,h_{i,k},$4, 100 epochs with cosine annealing, batch size $y_{i,k} = C\,h_{i,k},$5, gradient clipping with max-norm $y_{i,k} = C\,h_{i,k},$6, and cross-entropy loss with label smoothing. The train/test split is performed at breast level to avoid leakage (Bayatmakou et al., 23 Jul 2025).

The paper reports superior classification performance across all key metrics while maintaining computational efficiency. The summary performance values are:

Metric	Value
Accuracy	$y_{i,k} = C\,h_{i,k},$7
AUC	$y_{i,k} = C\,h_{i,k},$8
F1-score	$y_{i,k} = C\,h_{i,k},$9
Sensitivity	$h_{i,k}\in \mathbb{R}^{d_{HS}}$0
Specificity	$h_{i,k}\in \mathbb{R}^{d_{HS}}$1

For comparison, the best pure MambaVision-Large backbone is reported as $h_{i,k}\in \mathbb{R}^{d_{HS}}$2 in Accuracy/AUC, while a dual-stream MambaVision without SeqMoE is reported as $h_{i,k}\in \mathbb{R}^{d_{HS}}$3. The ablation study provides a more granular decomposition:

Variant	Accuracy	AUC
MambaVision on whole mammograms alone	0.7383	0.7748
MambaVision on cropped images	0.7517	0.7960
+ Dual-stream experts (no SeqMoE)	0.8322	0.8791
Full Mammo-Mamba (SeqMoE + dual-stream)	0.8792	0.9249

The paper interprets these gains as confirming three points: hybrid SSM/attention modeling provides efficiency without sacrificing global context, dual streams capture local versus global cues, and SeqMoE depth-wise gating yields additional precision by selectively refining challenging patches (Bayatmakou et al., 23 Jul 2025).

6. Relation to earlier mammography systems and Vision Mamba literature

Mammo-Mamba should be distinguished from the earlier system "MAMMO: A Deep Learning Solution for Facilitating Radiologist-Machine Collaboration in Breast Cancer Diagnosis" (Kyono et al., 2018). MAMMO is a clinical decision support system based on a multi-view convolutional neural network with multi-task learning and a separate triage network. On a private dataset of 8,162 patients, that system reduced the number of radiologist readings by 42.8% while improving the overall diagnostic accuracy in comparison to readings done by radiologists alone (Kyono et al., 2018). Its objective is radiologist–machine collaboration through triage, not the SecMamba–SeqMoE hybridization introduced in Mammo-Mamba.

Mammo-Mamba is also related to, but distinct from, "MedMamba: Vision Mamba for Medical Image Classification" (Yue et al., 2024). MedMamba introduced the SS-Conv-SSM hybrid block for generalized medical image classification and evaluated it on 16 datasets containing ten imaging modalities and 411,007 images. The MedMamba paper explicitly states that it did not include mammography. Its role here is therefore antecedent rather than task-specific: it established a Vision Mamba baseline in medical imaging, while Mammo-Mamba specializes hybrid state-space modeling for high-resolution multi-view mammography (Yue et al., 2024).

A common source of confusion is therefore nomenclature rather than methodology. “MAMMO” denotes a 2018 multi-task CNN plus triage framework, “MedMamba” denotes a generalized Vision Mamba classifier for medical images, and “Mammo-Mamba” denotes the 2025 hybrid MambaVision architecture with SecMamba and SeqMoE for multi-view mammogram classification. A plausible implication is that Mammo-Mamba occupies a specific point of convergence between mammography CAD research and the broader shift from purely attention-based vision models toward state-space–augmented architectures.