CMFDNet: Dual Roles in Medical Imaging

• The paper introduces CMFDNet for polyp segmentation, employing a Cross-Mamba Decoder, Multi-Scale Aware bridge, and Feature Discovery head to enhance boundary precision.
• CMFDNet also denotes a multimodal MRI framework that uses Complete Feature Disentanglement and Dynamic Mixture-of-Experts Fusion to achieve robust classification.
• Both implementations leverage specialized modules addressing scale variability, feature fusion, and latent disentanglement to improve overall diagnostic accuracy.

CMFDNet is an overloaded name in recent arXiv literature. In one usage, it denotes “CMFDNet: Cross-Mamba and Feature Discovery Network for Polyp Segmentation”, a Mamba/SSM-based medical image segmentation architecture for colonoscopy images that couples a Cross-Mamba Decoder (CMD), a Multi-Scale Aware (MSA) bridge, and a Feature Discovery (FD) head (Jiang et al., 25 Aug 2025). In another usage, the name is employed as a synonym for the complete network instantiating Completed Feature Disentanglement Learning (CFDL) for multimodal MRI classification, where the core components are Complete Feature Disentanglement (CFD) and Dynamic Mixture-of-Experts Fusion (DMF) (Liu et al., 2024). The shared acronym masks two unrelated methodological lineages: one is a 2D dense prediction model built around Vision Mamba and decoder-side feature fusion, while the other is a multimodal classification framework based on disentangled latent representations and expert gating.

1. Nomenclature and scope

The name CMFDNet is unambiguous only within the context of a given paper. In the polyp-segmentation paper, CMFDNet is the official model name and expands to Cross-Mamba and Feature Discovery Network (Jiang et al., 25 Aug 2025). In the multimodal MRI paper, the official method name is Completed Feature Disentanglement Learning, abbreviated CFDL; however, the provided overview explicitly states that CMFDNet refers to the complete network that implements CFD + DMF and is synonymous with the authors’ CFDL instantiation (Liu et al., 2024).

This naming overlap matters because the two systems differ in task, data modality, architectural primitives, and mathematical objectives. The segmentation CMFDNet operates on colonoscopy images resized to $224\times224$, uses a pretrained VMamba-Tiny encoder, and produces a pixelwise segmentation map through a top-down decoder (Jiang et al., 25 Aug 2025). The MRI-classification CMFDNet/CFDL uses non-shared 3D ResNet-18 backbones per modality, decouples latent vectors into shared, specific, and partial-shared components, and feeds a fused representation to an MLP classifier (Liu et al., 2024).

A common misconception is therefore to treat CMFDNet as a single canonical architecture. The literature provided does not support that view. Instead, it documents two separate architectures that happen to share the same shorthand.

2. CMFDNet as a polyp-segmentation architecture

In “CMFDNet: Cross-Mamba and Feature Discovery Network for Polyp Segmentation” (Jiang et al., 25 Aug 2025), the target problem is automated colonic polyp segmentation in colonoscopy. The paper frames the task around three failure modes: significant variation in polyp shapes and sizes, indistinct boundaries between polyps and adjacent tissues, and under-detection of small polyps because information is lost during downsampling and decoding. The architecture is explicitly designed against these failure modes.

The network uses VMamba-Tiny (Vision Mamba) as a pretrained encoder. Given an input image $I$ of size $H\times W$ with $H=W=224$, the encoder produces a three-stage feature pyramid $S_1,S_2,S_3$ at strides $\{4,8,16\}$, corresponding to spatial sizes $\{H/4,W/4\}$, $\{H/8,W/8\}$, and $\{H/16,W/16\}$ (Jiang et al., 25 Aug 2025). Each encoder feature is passed through an MSA module, and the resulting features feed a three-stage top-down CMD decoder. A downstream FD module then fuses decoder outputs across stages by multiplicative gating and residual addition, after which a final $1\times1$ convolution produces the segmentation map.

The architectural decomposition is as follows.

Component	Function	Key mechanism
MSA	Encoder-to-decoder bridge	Inverted bottleneck, multi-branch depthwise convolutions, ChannelShuffle
CMD	Decoder	Row/column exchange, diagonal SS2D cross-scanning, GAB, separable convolutions
FD	Prediction head fusion	Element-wise multiplication gating and residual fusion

The MSA module addresses scale diversity. It first applies GAB and pointwise convolution for channel expansion, then processes the feature map with parallel depthwise convolutions of sizes $I$0, $I$1, and $I$2, and finally aggregates them through Channel_Shuffle and a pointwise convolution: $I$3

$I$4

$I$5

The stated rationale is that polyps vary widely in apparent scale under different camera-to-tissue distances and endoscope optics (Jiang et al., 25 Aug 2025).

The CMD module is the decoder innovation. Its purpose is to reduce blurry boundaries by aligning deep semantic features with shallow local details through cross-scanning in complementary directions. At decoder stage $I$6, the current bridge feature $I$7 and the upsampled deeper decoder feature $I$8 undergo row and column exchange: $I$9

$H\times W$0

Each exchanged tensor is then processed by VMamba-style SS2D selective scanning in four diagonal orders $H\times W$1: $H\times W$2

$H\times W$3

The fused output is refined by GAB, pointwise convolutions, and depthwise convolution: $H\times W$4

$H\times W$5

The paper’s explanation is that row/column exchange emphasizes axial contrasts before diagonal SS2D, thereby increasing sensitivity to edge orientation and reducing boundary blur (Jiang et al., 25 Aug 2025).

The FD module targets small-polyp recovery. It introduces dependencies between decoder stages using transposed-convolution upsampling, element-wise multiplication, and residual addition: $H\times W$6

$H\times W$7

Here, $H\times W$8 denotes a transposed convolution with kernel $H\times W$9 and padding $H=W=224$0 that doubles spatial resolution. The multiplicative term acts as a per-pixel affinity or gate from deeper semantics onto finer features (Jiang et al., 25 Aug 2025).

3. Mathematical basis and training protocol of the segmentation model

The segmentation CMFDNet is grounded in the selective state-space model (SSM) formulation associated with Mamba. The paper states the general continuous-time linear SSM as

$H=W=224$1

and describes selective SSM as introducing data-dependent gating: $H=W=224$2

$H=W=224$3

In SS2D, 2D tensors are flattened into directional scan orders, and cross-scanning aggregates outputs from multiple directions. The broader stylized formulation provided includes eight directions,

$H=W=224$4

although this specific model uses four diagonal scans (Jiang et al., 25 Aug 2025).

The Global Attention Block (GAB) is defined as a parallel channel/spatial attention mechanism with learnable trade-off $H=W=224$5: $H=W=224$6

$H=W=224$7

The paper contrasts this with serial CBAM and states that weighted parallel fusion preserves both global channel importance and spatial saliency (Jiang et al., 25 Aug 2025).

Training uses deep supervision, but the paper does not specify the exact loss. The provided details recommend a standard compound loss matching the reported metrics, consisting of Dice loss and binary cross-entropy, summed over auxiliary side outputs and the final output. This is presented as a reproducibility recommendation rather than a claim that the paper explicitly used that exact formula (Jiang et al., 25 Aug 2025). The evaluation metrics explicitly reported are Dice, IoU, E-measure ($H=W=224$8), weighted F-measure ($H=W=224$9), structure measure ($S_1,S_2,S_3$0), and MAE.

The datasets and training pipeline are clearly specified. Training uses Kvasir-SEG (900 for training) and CVC-ClinicDB (550 for training); unused images from these datasets form visible test sets, while generalization is tested on ETIS-LaribDB, CVC-ColonDB, and EndoScene. Inputs are resized to $S_1,S_2,S_3$1, augmentation includes random horizontal flip, rotation, and color perturbations, and optimization uses AdamW with initial learning rate $S_1,S_2,S_3$2, halved every 50 epochs for a total of 150 epochs, batch size 8, under PyTorch on an NVIDIA A800 GPU (Jiang et al., 25 Aug 2025).

4. Empirical results, ablations, and limitations of the segmentation model

The segmentation CMFDNet is compared against PraNet, FCBFormer, ECTransNet, Swin-UMamba, VM-UNet, and VM-UNetV2. The paper states that it outperforms six SOTA methods, with the largest margins on the most difficult unseen datasets (Jiang et al., 25 Aug 2025).

The headline quantitative results are dataset-specific. On ETIS, the model reports mDice 81.85 and mIoU 74.27, with mDice exceeding the best SOTA baseline by 1.83%; it also records top $S_1,S_2,S_3$3 77.53, $S_1,S_2,S_3$4 94.12, $S_1,S_2,S_3$5 91.94, and MAE 1.35. On ColonDB, it reports mDice 83.05 and mIoU 75.56, improving the best SOTA mDice by 1.55%, with top $S_1,S_2,S_3$6 81.44, $S_1,S_2,S_3$7 88.01, $S_1,S_2,S_3$8 92.16, and MAE 2.99. On EndoScene, Kvasir, and ClinicDB, the reported mDice values are 90.62, 91.74, and 93.36, respectively, each with smaller positive margins over the best baseline (Jiang et al., 25 Aug 2025).

The ablation studies isolate the contributions of the main modules. On ColonDB, the full model yields mDice 83.05, mIoU 75.56, and MAE 2.99. Removing CMD reduces performance to mDice 77.12 and mIoU 70.31, with MAE rising to 3.67; removing MSA yields mDice 81.94 and mIoU 74.77; removing FD yields mDice 82.50 and mIoU 74.67; and replacing GAB with CBAM yields mDice 82.46 and mIoU 74.89. A parallel pattern is reported on ETIS, where removing CMD causes the largest drop, from 81.85/74.27 to 76.23/69.23 in mDice/mIoU (Jiang et al., 25 Aug 2025). The explicit conclusion given is that CMD delivers the largest gains, while MSA contributes multi-scale robustness, FD enhances small-polyp detection, and GAB improves attention fusion over CBAM in this design.

The paper also offers a state-of-the-art comparison by architectural family. It characterizes U-Net variants and CNN decoders such as PraNet as strong in boundary refinement but limited by local receptive fields and weaker global context modeling; hybrid CNN–Transformer models such as FCBFormer and ECTransNet as improving multi-scale semantics but incurring quadratic attention costs or relying on edge heuristics; and Mamba-based baselines such as VM-UNet, VM-UNetV2, and Swin-UMamba as lacking the particular combination of row/column exchange, diagonal multi-scan fusion, multi-branch MSA bridging, and FD gating used here (Jiang et al., 25 Aug 2025).

Several limitations are explicitly noted. The model does not contain dedicated noise suppression, even though colonoscopy commonly exhibits specular highlights and related artifacts. The paper further states that cross-dataset gains are strong, but larger domain shifts may require domain adaptation, style augmentation, or test-time adaptation. It also notes that ultra-small targets remain challenging, despite the FD module, and that directional SS2D contributions could be visualized to improve interpretability (Jiang et al., 25 Aug 2025). The reported complexity discussion is conceptual rather than empirical: parameter count, FLOPs, throughput, and FPS are not reported.

5. CMFDNet as Completed Feature Disentanglement Learning for multimodal MRI

In “Completed Feature Disentanglement Learning for Multimodal MRIs Analysis” (Liu et al., 2024), the overview states that CMFDNet refers to the complete network that implements CFD + DMF and is synonymous with the authors’ CFDL instantiation. This usage belongs to multimodal MRI classification rather than image segmentation.

The motivating problem is multimodal learning with more than two MRI sequences, such as T1, T1C/T1Gd, T2, FLAIR, DWI/ADC, and PD. The paper argues that conventional feature-disentanglement methods split modality representations into modality-shared and modality-specific components, then fuse them by concatenation or attention. When the number of modalities satisfies $S_1,S_2,S_3$9, this two-way decomposition can lose information shared only by a proper subset of modalities. The proposed remedy is Complete Feature Disentanglement (CFD), which augments shared and specific features with modality-partial-shared features (Liu et al., 2024).

Let $\{4,8,16\}$0 denote the number of modalities and let each modality-specific encoder $\{4,8,16\}$1 produce a latent vector $\{4,8,16\}$2 for sample $\{4,8,16\}$3. The paper defines:

• Modality-shared features

$\{4,8,16\}$4

• Modality-specific features

$\{4,8,16\}$5

• Modality-partial-shared features for subset $\{4,8,16\}$6 with $\{4,8,16\}$7

$\{4,8,16\}$8

The final set of decoupled features is

$\{4,8,16\}$9

Architecturally, the model comprises non-shared 3D ResNet-18 backbones per modality, single-FC disentanglement heads $\{H/4,W/4\}$0, $\{H/4,W/4\}$1, and $\{H/4,W/4\}$2, followed by a Dynamic Mixture-of-Experts Fusion (DMF) block. In three-modality settings, $\{H/4,W/4\}$3 contains all pairs; in four-modality settings, the model includes both pairs and triplets (Liu et al., 2024).

The disentanglement objective has three terms. First, all-modality shared similarity: $\{H/4,W/4\}$4 Second, within-subset partial-shared similarity: $\{H/4,W/4\}$5 Third, global dissimilarity across all final decoupled features: $\{H/4,W/4\}$6 The full objective is

$\{H/4,W/4\}$7

The paper emphasizes that this formulation uses MSE to pull shared and partial-shared features together and cosine similarity penalties to push distinct components apart, rather than mutual-information or adversarial constraints (Liu et al., 2024).

The DMF fusion computes a global context vector from the concatenation of all decoupled features and a local matrix from their stack. It then produces softmax expert weights: $\{H/4,W/4\}$8

$\{H/4,W/4\}$9

$\{H/8,W/8\}$0

$\{H/8,W/8\}$1

The fused representation $\{H/8,W/8\}$2 is passed to an MLP classifier (Liu et al., 2024).

6. Empirical profile of the MRI-classification model and the broader significance of the name overlap

The MRI-classification CMFDNet/CFDL is evaluated on three datasets: MRNet for binary meniscal tear, MEN for three-class meningioma grading/invasion, and BraTS 2021 for binary MGMT status classification (Liu et al., 2024). The training setup uses non-shared 3D ResNet-18 encoders, latent vectors of size 512, disentanglement heads with dim = 32, Adam with weight decay $\{H/8,W/8\}$3, batch size 32, linear warm-up for 5 epochs, step learning-rate decay by 0.8 every 5 epochs, and Dropout(0.5). Hardware is an NVIDIA RTX 3090 GPU under PyTorch (Liu et al., 2024).

Reported performance is strongest relative to baselines on multiple metrics. On MRNet, the model achieves the best ACC $\{H/8,W/8\}$4, G-Mean $\{H/8,W/8\}$5, Ba_ACC $\{H/8,W/8\}$6, AUPRC $\{H/8,W/8\}$7, and AUC $\{H/8,W/8\}$8. On MEN, it tops six metrics, including ACC $\{H/8,W/8\}$9, ACC_G2inv $\{H/16,W/16\}$0, ACC_G2ninv $\{H/16,W/16\}$1, weighted-F1 $\{H/16,W/16\}$2, macro-F1 $\{H/16,W/16\}$3, and AUC $\{H/16,W/16\}$4. On BraTS 2021, it leads in ACC $\{H/16,W/16\}$5, G-Mean $\{H/16,W/16\}$6, Ba_ACC $\{H/16,W/16\}$7, AUPRC $\{H/16,W/16\}$8, and AUC $\{H/16,W/16\}$9 (Liu et al., 2024). The paper further reports Wilcoxon signed-rank tests showing significant gains over many baselines on most metrics, particularly for MRNet and MEN.

Its ablations separate the roles of partial-shared disentanglement, mixture-of-experts, and local-in-global gating. On MRNet, enabling partial-shared disentanglement and adding LinG to MoE yields the best overall numbers, while naive MoE without LinG can hurt when the number of decoupled components increases (Liu et al., 2024). Heatmaps of gating weights show sample-wise and dataset-wise differences in which components dominate, such as strong weights for $1\times1$0 and $1\times1$1 in MRNet and for $1\times1$2, global shared $1\times1$3, and $1\times1$4 in MEN. The paper presents this as an interpretability advantage.

The model’s scalability is explicitly bounded. Enumerating all subsets is combinatorial, so the paper restricts $1\times1$5 to all pairs for $1\times1$6 and pairs plus triplets for $1\times1$7. Even under this restriction, the reported complexity remains competitive: 99.62M parameters and 46.16 GFLOPs for three modalities, and 132.89M parameters and 45.44 GFLOPs for four modalities (Liu et al., 2024). The paper also states that the current method assumes all modalities are present at train and test time; missing-modality handling is left for future work.

Taken together, the two CMFDNet usages illustrate a purely nominal convergence rather than a shared technical program. The segmentation model uses FD to mean Feature Discovery, whereas the multimodal MRI model uses FD to mean Feature Disentanglement in the broader discussion of prior work and introduces CFD as Complete Feature Disentanglement (Jiang et al., 25 Aug 2025, Liu et al., 2024). A plausible implication is that citations and reproductions should always resolve CMFDNet by paper title and task domain rather than acronym alone.