DyGLNet: Hybrid Medical Segmentation

• DyGLNet is a lightweight hybrid medical image segmentation network that fuses global attention with multi-scale dilated convolutions to tackle multi-scale lesion variation and blurred boundaries.
• It employs a symmetric U-shaped encoder-decoder architecture featuring the SHDCBlock for global-local feature extraction and DyFusionUp for dynamic, content-adaptive upsampling.
• The design bridges CNN and Transformer strengths to achieve superior boundary accuracy and small-object segmentation across diverse datasets while minimizing computational cost.

Searching arXiv for the cited DyGLNet paper and closely related dynamic-graph entries to ground the article in current arXiv records. DyGLNet is a lightweight hybrid medical image segmentation network proposed to address three recurring difficulties in clinical segmentation: multi-scale lesion variation, blurred or ill-defined boundaries, and the high computational cost of transformer-heavy architectures. It is formulated as a symmetric U-shaped encoder-decoder that fuses global and local features through a hybrid feature extraction module, SHDCBlock, and reconstructs feature maps through a dynamic adaptive upsampling module, DyFusionUp. The model is presented as an accurate and efficient architecture for polyp, brain tumor, dermoscopic lesion, and histopathological segmentation, with experiments on seven public datasets reporting strong results in boundary accuracy and small-object segmentation while maintaining relatively low computational complexity (Zhao et al., 16 Sep 2025).

1. Motivation and problem setting

DyGLNet is motivated by a specific mismatch between the requirements of medical image segmentation and the inductive biases of common backbone families. The underlying paper states that CNNs are efficient and strong at local texture and detail modeling, but have limited long-range modeling; Transformers capture global dependencies well, but are usually computationally expensive and weaker at local structure and boundary recovery; and standard upsampling methods such as bilinear interpolation blur edges, whereas transposed convolutions can introduce checkerboard artifacts and add cost (Zhao et al., 16 Sep 2025).

The target regime is characterized by lesions of very different sizes, irregular shapes, weak contrast with surrounding tissue, and tiny targets for which missing a few pixels can be consequential. DyGLNet is therefore designed around three coordinated responses: global context modeling with a lightweight attention branch, local detail modeling with multi-scale dilated convolutions, and dynamic, content-adaptive upsampling for sharper reconstruction. This design places DyGLNet between pure CNN and pure Transformer methods. A plausible implication is that its contribution is not merely hybridization in the generic sense, but a deliberately task-aligned decomposition of medical segmentation into long-range context, multi-scale local structure, and reconstruction fidelity.

2. Architectural organization

DyGLNet adopts a symmetric U-shaped encoder-decoder design. The encoder has four stages: the first stage uses basic convolution plus activation, and later stages use SHDCBlock-based feature extraction. The decoder has four upsampling stages, each of which uses DyFusionUp. Skip connections fuse encoder features at matching scales, and a final $1\times1$ convolution produces a single-channel segmentation map (Zhao et al., 16 Sep 2025).

The paper emphasizes a lightweight Transformer-enhanced convolutional network backbone with multi-level residual connections for full-scale feature fusion. In structural terms, the encoder is responsible for progressively extracting features that retain both long-range and local information, while the decoder reconstructs high-resolution segmentation outputs using dynamic upsampling rather than a fixed-grid rule. The use of matched-scale skip connections is consistent with the stated goal of preserving boundary-sensitive and small-structure information during reconstruction.

This organization is explicitly intended to balance accuracy and efficiency. The architectural choice is not described as a generic scaling-up of model capacity, but as a mechanism for coupling context aggregation and detail preservation under constrained computation.

3. SHDCBlock: hybrid global-local feature extraction

The principal encoder module is SHDCBlock, short for Single-Head Dilated Convolution block. Its function is to jointly model global relationships via attention and local context plus multi-scale shape cues via dilated depthwise convolutions (Zhao et al., 16 Sep 2025).

Given an input tensor

$X\in \mathbf{R}^{B\times C\times H\times W},$

the block splits channels into a global branch and a local branch:

$$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$

The channel division is identified in the paper as a key design choice and is later examined in ablation.

On the global path, SHDCBlock uses a lightweight single-head attention mechanism rather than standard multi-head attention. Before attention, the block applies DyT (Dynamic Tanh) instead of layer normalization:

$\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$

where $\alpha$, $\beta$, and $\gamma$ are learnable parameters. Query, key, and value tensors are produced through a $1\times1$ convolution:

$[Q,K,V]=Conv_{1\times1}(\phi(X_g)).$

Attention is then computed as

$$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}\left(\frac{Q^T K}{\sqrt{d}}\right)V.$$

The paper argues that single-head attention is cheaper, reduces memory and computation, is sufficient for modeling global dependencies in high-resolution medical images, and avoids overfitting to scale-specific patterns. DyT is argued to be advantageous because it does not rely on preset statistics, dynamically adjusts feature amplitude, and is more robust to large appearance changes, which is particularly relevant for MRI and CT where brightness and contrast vary greatly.

On the local path, SHDCBlock uses three depthwise convolutions with different dilation rates:

$X\in \mathbf{R}^{B\times C\times H\times W},$0

These outputs are fused with a residual connection and batch normalization:

$X\in \mathbf{R}^{B\times C\times H\times W},$1

The intended scale decomposition is explicit: $X\in \mathbf{R}^{B\times C\times H\times W},$2 captures local fine-grained structure, $X\in \mathbf{R}^{B\times C\times H\times W},$3 and $X\in \mathbf{R}^{B\times C\times H\times W},$4 provide broader context for larger or irregular lesions, the residual path preserves original local features, and batch normalization stabilizes the fused local output.

The two branches are concatenated and mixed by a $X\in \mathbf{R}^{B\times C\times H\times W},$5 convolution:

$X\in \mathbf{R}^{B\times C\times H\times W},$6

This fusion is described as providing channel mixing, dimensional adjustment, and nonlinear interaction between global and local features. The block therefore operationalizes the paper’s central claim that accurate segmentation requires collaborative modeling of long-range context and local structure rather than prioritizing one at the expense of the other.

4. DyFusionUp: dynamic adaptive upsampling and reconstruction

The decoder uses DyFusionUp, a dynamic upsampling module intended to reconstruct feature maps more faithfully than bilinear interpolation or transposed convolution. DyFusionUp is composed of dynamic upsampling, channel-alignment convolution, and a fusion-enhancement module, and is designed to improve boundary restoration, fine-detail reconstruction, and structural consistency (Zhao et al., 16 Sep 2025).

Upsampling begins by splitting each pixel into a $X\in \mathbf{R}^{B\times C\times H\times W},$7 subpixel grid with predefined offsets:

$X\in \mathbf{R}^{B\times C\times H\times W},$8

The paper notes that this is equivalent to standard bilinear interpolation when no learned displacement is applied. Given input features

$X\in \mathbf{R}^{B\times C\times H\times W},$9

the module predicts offsets with a $$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$0 convolution:

$$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$1

where $$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$2 is the number of channel groups. The predicted offsets are scaled as

$$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$3

which constrains the displacement range and stabilizes training. The refined sampling coordinate is

$$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$4

and the output at location $$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$5 is computed by interpolation:

$$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$6

This mechanism replaces fixed-grid upsampling with learned sampling positions. According to the paper, that adaptability is beneficial for irregular lesion boundaries, thin structures, low-contrast edges, and small objects requiring precise spatial placement. The sampling is performed in a group-wise manner to reduce computation.

After sampling, the upsampled feature map is aligned with skip-connected features by

$$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$7

The concatenated feature tensor

$$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$8

is then processed by a lightweight multi-scale dilated convolution module for cross-channel enhancement and context aggregation, followed by a final $$X=[X_g, X_l], \quad X_g\in \mathbf{R}^{B\times C_g\times H\times W}, \quad X_l\in \mathbf{R}^{B\times (C-C_g)\times H\times W}.$$9 convolution to fuse local spatial structure. In effect, DyFusionUp couples geometric adaptivity with lightweight feature fusion, and the paper repeatedly associates this with improved contour reconstruction and boundary accuracy.

5. Optimization, datasets, and evaluation protocol

DyGLNet is trained with a hybrid loss combining Dice loss and cross-entropy:

$\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$0

$\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$1

and the final objective is

$\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$2

with

$\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$3

The paper states that this balances region overlap, pixel-level boundary accuracy, and class imbalance robustness (Zhao et al., 16 Sep 2025).

Implementation details are specified as follows: PyTorch framework, NVIDIA RTX 4090 GPU, input size $\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$4, AdamW optimizer, betas $\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$5, initial learning rate $\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$6, weight decay $\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$7, polynomial decay learning-rate schedule, warm-up of 10 epochs, 130 total epochs, batch size 16, mixed precision enabled, random seed 42, and gradient clipping. Data augmentation includes random cropping, horizontal flip, vertical flip, random rotation in $\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$8, elastic deformation, and brightness/contrast perturbation. Evaluation uses Dice, IoU, Precision, Recall, Specificity, and Accuracy.

The seven public datasets used in evaluation are summarized below.

Dataset	Image type	Split
Kvasir-SEG	Colon polyp images	800/100/100
CVC-ClinicDB	Colonoscopy frames	489/61/62
Brain-MRI	Tumor MRI images	1089/137/138
ISIC2016	Dermoscopic images	900 train, 379 test
PH2	Dermoscopic images	160/20/20
GlaS	Colorectal histology images	132/16/17
TNBC	Breast cancer tissue images	40/10

The breadth of datasets is significant because it spans polyp segmentation, brain tumor segmentation, dermoscopic lesion segmentation, and pathological tissue segmentation. This suggests that the model is evaluated across multiple imaging modalities and multiple regimes of scale variation and boundary ambiguity rather than a single benchmark niche.

6. Empirical performance, efficiency, and ablation

The paper reports that DyGLNet outperforms existing methods on seven public datasets and particularly excels in boundary accuracy and small-object segmentation (Zhao et al., 16 Sep 2025). On Kvasir-SEG, it achieves Dice $\phi(X_g)=\gamma\cdot \tanh(\alpha X_g)+\beta,$9 and IoU $\alpha$0, exceeding CMUNeXt-L by $\alpha$1 Dice and $\alpha$2 IoU, and surpassing FATNet, identified as the strongest hybrid baseline in that table. On CVC-ClinicDB, it reaches Dice $\alpha$3 and IoU $\alpha$4, described as the best result in the comparison set. The paper specifically notes strength on blurred polyp borders and tiny polyps, including cases with diameter $\alpha$5 mm.

For Brain-MRI, DyGLNet achieves Dice $\alpha$6, IoU $\alpha$7, Precision $\alpha$8, Recall $\alpha$9, Specificity $\beta$0, and Accuracy $\beta$1. The paper highlights that this yields the best Dice and IoU, the best Precision and Specificity, and very strong Recall. For dermoscopic lesion segmentation, DyGLNet attains Dice $\beta$2, IoU $\beta$3, and Accuracy $\beta$4 on ISIC2016, and Dice $\beta$5, IoU $\beta$6, Precision $\beta$7, and Accuracy $\beta$8 on PH2. It is described as especially strong for irregular melanoma shapes, early small lesions, and avoiding confusion with surrounding pigmentation. In histopathology, it reports Dice $\beta$9, IoU $\gamma$0, Precision $\gamma$1, and Accuracy $\gamma$2 on GlaS, and Dice $\gamma$3, IoU $\gamma$4, Precision $\gamma$5, and Accuracy $\gamma$6 on TNBC, with the paper stressing that it generalizes well despite the TNBC dataset having only 50 images.

Efficiency is a central component of the claimed contribution. The reported model complexity is 9.98M parameters and 11.16G FLOPs. Comparative figures given in the paper include H2Former at 33.63M parameters and 24.67G FLOPs, FATNet at 29.62M parameters and 42.80G FLOPs, and BEFUnet at 42.61M parameters and 7.95G FLOPs. The stated computational advantages are attributed to single-head instead of multi-head attention, depthwise dilated convolutions, group-wise dynamic upsampling, and lightweight $\gamma$7 and $\gamma$8 convolutions.

The ablation study supports the necessity of each major component. Performance drops on all datasets without DyT; replacing DyFusionUp with BiIn or TConv is worse; removing dilated convolutions degrades results, especially on Brain-MRI; removing single-head attention also hurts performance; and altered channel allocations in Split-DC and Split-SA worsen results. The reported interpretation is that DyT improves adaptability to modality variations, dynamic upsampling is better than fixed-grid or transposed upsampling, both local and global branches are necessary, and balanced channel division is important for branch cooperation. The full DyGLNet is concluded to be the best configuration across all seven datasets.

7. Nomenclature and relation to dynamic-graph research

DyGLNet belongs to medical image segmentation, not to the dynamic-graph architectures that have similar acronyms. The similarity in naming can obscure that these are distinct research lines. In particular, the paper titled "Dynamic Graph: Learning Instance-aware Connectivity for Neural Networks" proposes DG-Net, a network initialized as a complete directed acyclic graph whose routers learn sample-dependent edge weights and select active subgraphs for image classification and object detection; this is a dynamic connectivity mechanism for CNN backbones such as MobileNetV2, ResNet, ResNeXt, and RegNet, rather than a U-shaped medical segmentation model (Yuan et al., 2020).

A different line of work concerns Dynamic Graph Neural Networks on evolving graphs. "Towards Dynamic Graph Neural Networks with Provably High-Order Expressive Power" studies DyGNNs on continuous-time dynamic graphs, introduces $\gamma$9-dimensional Dynamic WL tests, proves that existing DyGNNs are upper bounded by the 1-DWL test, and proposes HopeDGN, a node-pair-level model with expressive power equivalent to the 2-DWL test under injectivity assumptions (Wang et al., 2024). That paper explicitly states that its experiments are not specifically on a model named DyGLNet.

Accordingly, DyGLNet in the strict sense of (Zhao et al., 16 Sep 2025) should be understood as a lightweight hybrid segmentation network centered on SHDCBlock and DyFusionUp. DG-Net (Yuan et al., 2020) and HopeDGN (Wang et al., 2024) are relevant mainly as terminological contrasts: they share a surface association with "dynamic" and "graph" terminology, but they address different data types, different mathematical objects, and different learning problems.