MGF-Skip: Mamba-Guided Fusion Skip Connection

• The paper demonstrates that MGF-Skip enhances encoder-decoder segmentation by using decoder-based gating to suppress noise and improve feature integration.
• MGF-Skip employs gated convolutions, residual reinforcement, and concatenation, enabling efficient boundary localization in medical image segmentation.
• Empirical results reveal that MGF-Skip boosts performance metrics like IoU and DSC compared to traditional concatenation and attention modules.

The Mamba-Guided Fusion Skip Connection (MGF-Skip) is a skip connection module designed to enhance semantic and spatial feature integration in encoder–decoder architectures for medical image segmentation. Introduced in HyM-UNet, MGF-Skip leverages semantically rich decoder features as gating signals to suppress noise in encoder features while enforcing fine structural detail through a residual pathway. This configuration enables improved boundary localization and noise robustness compared to conventional concatenation or attention-based skips, directly addressing the semantic gap and feature misalignment common in deep convolutional architectures (Chen et al., 22 Nov 2025).

1. Architectural Placement and Input–Output Relations

MGF-Skip replaces the standard feature concatenation employed at each encoder–decoder interface of U-Net-derived segmentation models. At a given decoder stage $i$, the module receives:

• $E_i \in \mathbb{R}^{H_i \times W_i \times C_e}$: the spatially high-resolution, texture-rich feature map from encoder stage $i$.
• $$D_i \in \mathbb{R}^{H_{i+1} \times W_{i+1} \times C_d}$$: a lower-resolution, semantically strong decoder feature from the previous decoding stage.

$D_i$ undergoes upsampling, $$D_i^{up} = \mathrm{Upsample}(D_i) \in \mathbb{R}^{H_i \times W_i \times C_d}$$, aligning spatial dimensions. The MGF-Skip module fuses $E_i$ and $D_i^{up}$ to produce $$F_{skip} \in \mathbb{R}^{H_i \times W_i \times (C_e + C_d)}$$, which is then consumed by the subsequent decoder block. Fig. 1 of (Chen et al., 22 Nov 2025) visually delineates this interface in the overall architecture.

2. Mathematical Formulation

The fusion process in MGF-Skip involves spatial gating, feature filtering, residual reinforcement, and concatenation:

1. Gate Computation: The upsampled decoder feature $D_i^{up}$ undergoes a sequence of convolutions and nonlinearities to derive a spatially-aware gate:

$$G_i = \sigma \left( \mathrm{Conv}_{1\times1} \left( \mathrm{ReLU} \left( \mathrm{Conv}_{3\times3}(D_i^{up}) \right) \right) \right) \in [0,1]^{H_i \times W_i \times C_e}$$

• $\mathrm{Conv}_{3\times3}$ maps $C_d \rightarrow C_e$ channels, stride 1, padding 1.
• $\mathrm{Conv}_{1\times1}$ maps $C_e \rightarrow C_e$.
• $\sigma(\cdot)$ denotes the sigmoid activation.

2. Feature Filtering: The encoder features are modulated spatially:

$E_{filt} = E_i \odot G_i$

where $\odot$ denotes element-wise multiplication.

3. Residual Reinforcement: The gated and original encoder features are summed:

$E_{res} = E_i + E_{filt}$

4. Final Fusion: The residual-augmented encoder feature is concatenated with the upsampled decoder feature:

$F_{skip} = \mathrm{Concat}[E_{res},\; D_i^{up}]$

This fused feature serves as input to the next decoder stage.

3. Implementation Specifics

MGF-Skip’s gating branch comprises:

• $3\times3$ convolution ($C_d \rightarrow C_e$, stride 1, padding 1)
• ReLU activation
• $1\times1$ convolution ($C_e \rightarrow C_e$)
• Sigmoid activation

Batch normalization is intentionally omitted in the gating branch to retain spatial sensitivity. Channel dimensions and kernel sizes are configured to ensure alignment between the gating mask (%%%%24%%%%) and the encoder feature ($E_i$). The residual addition $E_i + E_{filt}$ constitutes an internal skip within the module, ensuring information preservation.

4. Fusion Strategy and Functional Rationale

MGF-Skip is architected to achieve two critical objectives:

• Suppression of Background Noise: The sigmoid-activated gating mask $G_i$, derived from deep decoder features, adaptively down-weights locations in $E_i$ associated with image noise or irrelevant structures (e.g., artifacts from hair occlusion or specular highlights). This mechanism is fully differentiable and trained end-to-end, with no fixed thresholds or hard masking.
• Preservation and Enhancement of Boundaries: Given that aggressive gating may suppress true boundary information, the residual connection $E_{res} = E_i + E_{filt}$ ensures that low-level spatial details, essential for precise contour delineation, are continuously reinforced.

All convolutional parameters within the gating branch are learnable and jointly optimized with the rest of HyM-UNet.

5. Integration within HyM-UNet and Training Configuration

MGF-Skip is instantiated at each of the four encoder–decoder transition stages. Inputs are drawn from a hybrid encoder: early stages utilize CNN blocks for local texture modeling, while deeper stages employ Visual Mamba modules for long-range context. The training protocol for HyM-UNet with MGF-Skip includes:

• Optimizer: AdamW (weight decay $=10^{-4}$, $\beta_1=0.9$, $\beta_2=0.999$)
• Initial learning rate: $10^{-4}$, cosine-annealed to $10^{-6}$ over 200 epochs
• Batch size: 24
• Input resolution: $256 \times 256$
• Loss:

$$\mathcal{L}_{total} = \mathcal{L}_{Dice} + 0.5\,\mathcal{L}_{BCE} + 0.5\,\mathcal{L}_{Edge}$$

combining Dice, binary cross-entropy, and boundary-aware edge loss.

6. Empirical Performance and Comparative Analysis

Table 1 of (Chen et al., 22 Nov 2025) demonstrates that HyM-UNet incorporating MGF-Skip achieves superior results on the ISIC 2018 test set:

• Intersection over Union (IoU): $81.82\%$
• Dice Similarity Coefficient (DSC): $88.97\%$
• 95th percentile Hausdorff Distance (HD95): $4.03$ mm
• Precision: $90.91\%$

These scores surpass those of U-Net, CE-Net, and Attention U-Net. Table 2 presents an ablation: integrating MGF-Skip into a U-Net baseline increases IoU by $+1.02\%$ (from $80.13\%$ to $81.15\%$) and DSC by $+0.31\%$ (from $87.81\%$ to $88.12\%$), outperforming SE and CBAM attention modules in the same positions.

MGF-Skip's design results in minimal additional parameter overhead relative to standard skip connections or spatial/channel attention modules, while maintaining inference latency ($\approx 12$ ms per $256 \times 256$ image on RTX 3090) competitive with or lower than ViT-based architectures.

7. Distinctions from Prior Skip Connection Mechanisms

MGF-Skip contrasts with alternative strategies as follows:

Skip Module	Gating Source	Residual Path	Attention Dimension
Standard concat	None	None	None
SE	Encoder/Decoder	None	Channel
CBAM	Encoder/Decoder	None	Channel + Spatial
MGF-Skip	Decoder (deep)	Encoder	Spatial (decoder-guided)

Standard concatenation indiscriminately propagates all encoder features, including noise. SE and CBAM introduce channel/spatial attention but lack explicit residual renormalization and do not leverage decoder semantics for gating. MGF-Skip’s use of the decoder as a gating source, with end-to-end-learned dynamic suppression and explicit residual addition, is unique in enhancing ambiguous boundary regions and suppressing artifacts, as shown by empirical evaluation (Chen et al., 22 Nov 2025).