DAUNet: Advanced U-Net for Segmentation

• DAUNet is a family of U-Net variants incorporating dual-attention and deformable convolution modules to improve segmentation across various medical imaging modalities.
• The 3D variant uses dual-attention at decoder skip connections for volumetric MRI segmentation, while the 2D lightweight model employs SimAM attention for CT and ultrasound imaging.
• Both architectures deliver higher accuracy and efficiency, achieving improved Dice scores and robustness compared to traditional U-Net models.

DAUNet refers to two independently developed U-Net variants integrating advanced attention and adaptability modules for medical image segmentation. Both architectures—(1) the 3D Dual-Attention U-Net for volumetric meningioma segmentation in MRI (Bouget et al., 2021) and (2) the 2D lightweight DAUNet utilizing Deformable V2 convolutions with SimAM parameter-free attention (Munir et al., 7 Dec 2025)—leverage attention-driven feature fusion but diverge in their mechanisms, structural details, and application domains.

1. Definitions and Architectural Overview

DAUNet denotes distinct, attention-enriched extensions of U-Net, aimed at enhancing the segmentation performance on heterogeneous, complex medical images.

• In the context of T1-weighted MRI meningioma segmentation (Bouget et al., 2021), DAUNet is a 3D U-Net backbone augmented with Dual-Attention Modules at every decoder skip connection, multi-scale input pathways into the encoder, and deep supervision applied to all decoder outputs. The dual-attention mechanism fuses spatial and channel self-attention, improving context capture and focus on salient features.
• As a lightweight segmentation model for ultrasound and CT (Munir et al., 7 Dec 2025), DAUNet features a standard U-Net encoder–decoder macro-architecture but replaces the bottleneck layer with a Deformable V2 convolution block followed by insertion of SimAM parameter-free attention. All skip connections undergo SimAM before concatenation, enhancing saliency while maintaining parameter efficiency.

Both designs retain the hallmark encoder–decoder structure of U-Net for multiscale feature aggregation but differ in their implementation of attention and adaptability to spatial and contextual heterogeneity.

2. Key Components and Modules

Dual-Attention Module (3D DAUNet, (Bouget et al., 2021))

The Dual-Attention Module (DAM) simultaneously models channel and spatial dependencies:

• Channel Attention: For feature tensor $\mathbf{F}\in\mathbb{R}^{C\times N}$, computes similarity $$S_c = \operatorname{Softmax}(\mathbf{F}_c \mathbf{F}_c^\top)$$ over channels, yielding attended features $\mathbf{F}_c'=S_c \mathbf{F}_c$.
• Spatial Attention: Computes spatial map $$S_s = \operatorname{Softmax}(\mathbf{F}_s \mathbf{F}_s^\top)$$, then $\mathbf{F}_s' = \mathbf{F}_s S_s$.

Both branches are projected via $1\times1\times1$ convolutions and summed with the original feature:

$$\hat{\mathbf{F}} = \mathbf{F} + W_c(\mathbf{F}_c') + W_s(\mathbf{F}_s')$$

DAMs are placed prior to skip connection concatenation at every decoding level, following joint embedding of encoder and decoder feature maps.

Deformable V2 Convolution and SimAM (Lightweight DAUNet, (Munir et al., 7 Dec 2025))

• Deformable V2 Convolution: Enhances the bottleneck by allowing each kernel position $p_n$ to have a learnable offset $\Delta p_n$ and modulation $\alpha_n \in [0,1]$:

$$y(p_0) = \sum_{p_n\in\mathcal R} \alpha_n w(p_n) \cdot x(p_0 + p_n + \Delta p_n)$$

Offsets and modulation scalars are inferred via auxiliary convolutional layers and optimized end-to-end, giving the model spatial sampling flexibility.

• SimAM Attention: Injected post-bottleneck and along all skip pathways, SimAM assigns a saliency score to each neuron $t$ in a channel by estimating energy:

$$E_t = (x_t - \mu_t)^2 + \lambda \sum_{i\neq t}(x_i - \mu_t)^2$$

where $\mu_t$ is the mean over the rest of the channel, and the attention mask is $a_t = \sigma(1/(E_t + \epsilon))$. Output features are filter-wise rescaled without added parameters.

3. Network Workflow and Data Flow

3D DAUNet: MRI Tumor Segmentation

• Encoder: Five hierarchical levels with paired 3D convolutions; multi-scale downsampled inputs are injected at each level to preserve spatial frequency information.
• Decoder: Each level upsamples and fuses the corresponding skip connection (post-dual-attention), with deep supervision via 1×1×1 convolutions at each level.
• Supervision: Multi-level (deep supervision) losses averaged over decoder outputs; the main loss function is the soft Dice loss.

Lightweight DAUNet: Ultrasound/CT Segmentation

• Encoder: Four levels of 2D convolution + ReLU, downsampled by maxpooling.
• Bottleneck: Channel-compressed by $1\times1$ convolution, processed by a 3×3 Deformable V2 convolution, then restored in channel count; SimAM attention module follows.
• Decoder: For each level, upsample 2×, apply SimAM to encoder skip, concatenate, process via dual convolutional layers, and repeat.
• Output: Final $1 \times 1$ convolution maps features to the segmentation mask.

Pseudocode for a forward pass (single image):

skips = [] h = x for level in [1..4]: h = conv3x3_ReLU(h) h = conv3x3_ReLU(h) skips.append(h) h = maxpool2x2(h) h = conv1x1_compress(h) h = deformable_conv3x3(h) # Deformable V2 Conv h = conv1x1_expand(h) h = SimAM(h) for level in [4..1]: h = upsample2x(h) skip_feat = skips[level] skip_feat = SimAM(skip_feat) h = concat(h, skip_feat) h = conv3x3_ReLU(h) h = conv3x3_ReLU(h) output = conv1x1(h) return output

4. Evaluation Protocols and Results

Datasets

• MRI meningioma segmentation (Bouget et al., 2021): 600 Gd-T1 volumes, resampled to $128^3$; 5-fold cross validation.
• Fetal head and pubic symphysis ultrasound & FUMPE CT embolism detection (Munir et al., 7 Dec 2025):
  • FH-PS-AoP: 4,000 train, 700 test images; all resized to $256\times256$.
  • FUMPE: 8,792 slices, 3,438 annotated regions.

Metrics

• Dice Similarity Coefficient (DSC):

$\mathrm{DSC}(P,G) = \frac{2|P\cap G|}{|P| + |G|}$

• Hausdorff Distance 95 (HD95)
• Average Symmetric Surface Distance (ASD)

Performance Summary

Model	Application	DSC (%)	HD95	ASD	Parameters (M)
DAUNet (Munir et al., 7 Dec 2025)	FH-PS-AoP (US, 3 classes)	89.09/92.87/85.31	10.37/12.76/7.98	3.70/4.44/2.95	20.47
UNet	FH-PS-AoP	80.22/81.42/79.02	15.87/17.42/14.33	4.88/5.09/4.67	31.03
TransUNet	FH-PS-AoP	87.34/93.02/81.66	—	—	105.28
DAUNet (Munir et al., 7 Dec 2025)	FUMPE (CT)	88.80	2.57	—	20.47
SCUNet++	FUMPE (CT)	83.47	3.83	—	60.11

On Gd-T1 MRI meningiomas (Bouget et al., 2021), DAUNet achieves a patient-wise Dice of $78.43\pm27.56\%$, F1-score $92.99\pm2.76\%$, recall $91.19\pm4.71\%$, and precision $95.04\pm2.89\%$; outperforming UNet-FV and AGUNet-AG by 2-3% absolute Dice.

Ablation studies in both frameworks highlight significant and additive gains of attention and deformable modules.

Configuration (Deformable?, SimAM?)	DSC	HD95	ASD	Params (M)
Baseline UNet (×,×)	80.22	15.87	4.88	31.03
+SimAM only (×,✓)	82.16	14.73	4.92	31.03
+Deformable only (✓,×)	87.67	11.87	4.09	20.47
Full DAUNet (✓,✓)	89.09	10.37	3.70	20.47

5. Analysis of Robustness, Efficiency, and Deployment Suitability

• Parameter Efficiency: Lightweight DAUNet attains higher segmentation accuracy at significantly reduced model size (20.47 M parameters) compared to transformer-based (105.28 M) and other multi-branch networks, due to channel compression and SimAM's zero-parameter design.
• Robustness: Deformable V2 layers in the bottleneck maintain structured sampling even with partial or masked context, enabling accurate segmentation under missing data—a critical property for clinical edge deployment.
• Inference and Resource Requirements: The architectures achieve rapid inference (reported $\sim$3.1 s/volume on GPU for 3D DAUNet (Bouget et al., 2021)), and the lightweight version is suitable for real-time, memory-constrained settings.

6. Limitations and Future Directions

Both DAUNet variants present constraints related to input resolution, representation of very small structures, and potential under-utilization of available multimodal data. In T1-weighted MRI, tumors $<3$ ml and non-enhancing regions are sometimes missed; simple attention mechanisms yielded modest improvements unless combined with deep supervision or multi-scale strategies.

Suggested enhancements include:

• Integration of larger numbers of small lesion cases and multi-modal MR sequences.
• Incorporation of spatial-pyramid or atrous-spatial-pyramid context modules to supplement multi-scale fusion.
• Instance-aware cascade architectures or joint segmentation–detection heads tuned for small-object sensitivity.
• Loss function adaptation to penalize small-object missegmentations.

A plausible implication is that future DAUNet architectures may further unify parameter-free attention and deformable sampling into fully 3D implementations, particularly as computational resources for high-resolution volumetric imaging improve.

7. Significance in Medical Image Segmentation

DAUNet, in both its manifestations, represents a general approach of integrating advanced attention and spatially adaptive components into encoder–decoder architectures. It demonstrates practically meaningful improvements in segmentation for challenging domains (e.g., ultrasound, CT, and MRI brain tumor data), delivering higher precision, robustness, and efficiency than prior baselines. The frameworks provide reproducible benchmarks for further research into lightweight and context-aware medical image segmentation methods.

(Munir et al., 7 Dec 2025, Bouget et al., 2021)