Dual-Domain/Branch U-Nets

• Dual-Domain/Branch U-Nets are advanced neural network architectures that split processing into distinct branches tailored to separate data domains or semantic tasks.
• They employ mechanisms like bidirectional enhancement and learned scalar fusion to enforce consistency and optimally merge complementary features.
• These architectures achieve improved segmentation and image reconstruction performance by integrating local and global cues across different imaging modalities.

Dual-domain and dual-branch U-Nets generalize the classic U-Net architecture by introducing explicit architectural bifurcation—either across distinct spatial-spectral/physical domains (dual-domain), or across different semantic or modality-specific branches (dual-branch). These modifications are designed to jointly leverage complementary cues (e.g., local-vs-global, body-vs-boundary, image-vs-k-space, or multimodal representations) in a structured, often symmetrically coordinated manner, producing empirically measurable gains across medical imaging applications. Key themes include separate but interactive encoding/decoding paths, domain- or task-specific parameterization, and dedicated mechanisms for information exchange or consistency enforcement between branches.

1. Dual-Domain and Dual-Branch U-Net Fundamentals

Dual-domain or dual-branch U-Nets augment the encoder–decoder (“U”) architecture of the classic U-Net by maintaining two or more parallel computational streams, typically realized as separate branches within the network, each specialized either for a physical domain (e.g., image and Fourier/k-space), a semantic target (e.g., body and boundary regions), a modality (e.g., CT and MRI), or a feature type (e.g., convolutional and Kolmogorov–Arnold nonlinear layers).

• Dual-domain usually refers to architectures with branches operating explicitly in different data representations, such as image domain and frequency (k-space/spectral) domain (Souza et al., 2019, Liu et al., 2022, Farshad et al., 2022).
• Dual-branch commonly describes networks with functional split by semantic or substructure categories (e.g., region-vs-boundary), by modality, or by feature extractor type (Xu et al., 2024, Fang et al., 2024, Qu et al., 4 Aug 2025).

The core architectural motivation is to enhance representational capacity and information integration beyond what a monolithic single-branch network can achieve, while preserving the strengths of symmetrical encoder–decoder processing and skip connections.

2. Principal Architectural Variants

2.1 Body–Boundary Dual-Branch (DBF-Net Style)

DBF-Net exemplifies the semantic dual-branch paradigm, using a single encoder funneling into two decoders:

• Body branch targets interior (region) segmentation.
• Boundary branch delineates fine edge structures.

Feature interactions are implemented via parallel convolutional pathways within feature fusion/supervision (FFS) blocks, followed by bidirectional enhancement: $$F^{*}_\text{body,i} = F_{\text{body},i} + \operatorname{Conv}_{3\times3}(F_{\text{bound},i}), \qquad F^{*}_\text{bound,i} = F_{\text{bound},i} + \operatorname{Conv}_{3\times3}(F_{\text{body},i})$$ Outputs are adaptively merged through a learned scalar weight $\lambda$ (Xu et al., 2024).

2.2 Image–k-Space (Dual Domain) Cascades

In MR image reconstruction, dual-domain cascades such as the W-net and KV-Net alternate or parallelize U-Nets for image-domain and k-space-domain processing (Souza et al., 2019, Liu et al., 2022):

• In alternating cascades (e.g., W-net IK or KI), image and k-space domain U-Net blocks process outputs sequentially, with hard data consistency imposed at measured k-space samples.
• In parallel-fusion cascades (e.g., KV-Net), image- and k-space-specific sub-networks (V-Net and K-Net) process in parallel, with outputs fused using a learned parameter $\mu$ at each cascade stage: $I^{(t)} = \frac{A^{(t)}_i + \mu A^{(t)}_k}{1+\mu}$ This enables simultaneous integration of local (image) and global (spectral) corrections (Liu et al., 2022).

2.3 Spatiospectral (Spatial–Frequency) Dual-Encoder

Y-Net combines spatial and spectral (Fourier) feature encoding:

• Spatial encoder: standard U-Net downsampling path.
• Spectral encoder: Fast Fourier Convolution (FFC) blocks with Fourier-domain transforms, non-local mixing, and reweighting.
• Features are fused at the bottleneck and passed to a shared decoder (Farshad et al., 2022).

2.4 Multi-domain/Task Adapters (3D U²-Net)

Here, “dual-domain” refers to task or dataset domains. Each convolutional layer decomposes into a domain-specific depthwise convolution followed by a shared pointwise convolution: $$\text{Output} = W_\text{pointwise} * (W_{\text{depthwise}}^{(t)} * \text{Input})$$ Multiple tasks share the core network, with only lightweight domain adapters learned per task (Huang et al., 2019).

2.5 Heterogeneous Feature Extractors (KAN-Convolution Dual Channel)

KANDU-Net processes features via both conventional convolutional U-Net branches and per-pixel KAN (Kolmogorov–Arnold Network) nonlinear layers, fusing their outputs at each block using an auxiliary learned network (Fang et al., 2024).

2.6 Dual-Modality (CT/MRI) Alignment and Fusion

RL-U²Net uses separate Swin Transformer–based encoder–decoders for each modality, coordinated through reinforcement learning–guided cross-modal feature alignment (RL-XAlign). The aligned representations are decoded independently, then ensembled in the final segmentation (Qu et al., 4 Aug 2025).

3. Feature Fusion and Consistency Mechanisms

Branch and domain outputs are merged using methods that enforce mutual consistency and balance:

• Learned Scalar Fusion: In DBF-Net and KV-Net, scalar weights ($\lambda$ or $\mu$) learn to optimize the trade-off between branch contributions (Xu et al., 2024, Liu et al., 2022).
• Bidirectional Enhancement: Branches enhance each other via cross-convolutions before merging (Xu et al., 2024).
• Domain-Consistency Enforcement: Explicit “data consistency” is imposed in dual-domain MRI architectures by correcting predicted k-space or image-domain values to match measured samples (Souza et al., 2019, Liu et al., 2022).
• Auxiliary Fusion Networks: In KANDU-Net, a dedicated network fuses convolutional and KAN branch outputs (Fang et al., 2024). In RL-U²Net, fusion is guided by reinforcement learning to align cross-modal features (Qu et al., 4 Aug 2025).
• Skip Connections: Most architectures maintain U-Net style skip connections using either one branch’s features (Y-Net (Farshad et al., 2022)) or fused features (KANDU-Net (Fang et al., 2024)).

4. Training Objectives and Supervision

Multi-branch approaches frequently employ multi-task supervision. Examples include:

• DBF-Net: Combined loss over final segmentation, intermediate body, and boundary outputs. Each supervised via a mixed weighted binary cross-entropy and Dice loss, with customized pixel reweighting for sparse foregrounds (Xu et al., 2024).
• Dual-domain MRI cascades: Use mean squared error or SSIM-like losses on reconstructed images (Souza et al., 2019, Liu et al., 2022).
• KANDU-Net: Optimizes both cross-entropy and auxiliary Dice loss, with different learning rates for main and fusion components (Fang et al., 2024).
• RL-U²Net: Main segmentation losses are adaptively balanced (AGWD scheme), with auxiliary losses for alignment and RL policy/value. Segmentation and alignment losses are combined with PPO-based updates for the RL agent (Qu et al., 4 Aug 2025).
• Y-Net: Standard Dice and cross-entropy loss functions at the semantic segmentation output (Farshad et al., 2022).
• 3D U²-Net: Hybrid Lovász-Softmax and focal losses, each computed per-domain (Huang et al., 2019).

5. Empirical Results and Application Domains

The dual-domain/branch U-Net paradigm delivers consistently superior or at least competitive results over single-branch baselines across a range of applications:

Architecture	Application	Key Metric(s) (Test)	Benchmark Improvement	Reference
DBF-Net	Ultrasound lesion segmentation	Dice 81.05% (BUSI), 76.41% (UNS), 87.75% (UHES)	Outperforms U-Net, DeepLabV3+, LinkNet, UNeXt	(Xu et al., 2024)
W-net, KV-Net	MR image reconstruction (multi-coil)	SSIM 0.7814, NMSE 0.0271 (fastMRI test)	Matches i-RIM/XPDNet at 10× fewer parameters (SSIM gain over U-Net)	(Souza et al., 2019); (Liu et al., 2022)
Y-Net	OCT segmentation	Fluid Dice 0.93 (+13% rel. vs U-Net)	Average Dice gain 1.9%	(Farshad et al., 2022)
3D U²-Net	Multi-organ, multi-domain segmentation	Mean Dice ≈83.1%	1% overall param count, matched accuracy to per-task U-Nets	(Huang et al., 2019)
KANDU-Net	Nucleus/gland/US tumor segmentation	DSC: 94.1% (MoNuSeg), F1: 93.6% (GLAS)	Exceeds U-Net, U-Net++, U-KAN, U-Mamba	(Fang et al., 2024)
RL-U²Net	3D whole-heart segmentation (CT/MRI)	Dice: 93.1% (CT), 87.0% (MRI)	SOTA on MM-WHS 2017, sharpest cross-modality consistency	(Qu et al., 4 Aug 2025)

In summary, across imaging domains—ultrasound, MRI, OCT, histology, multimodality—these architectures consistently realize improved edge delineation, fidelity, and adaptation efficiency compared to their monolithic U-Net counterparts.

6. Design Trade-offs and Ablative Insights

• Branch Design: The optimal choice of branch specialization depends on the nature of the representation gap (e.g., semantic, physical, or modality). For example, in multi-coil MRI, pure image-domain networks suffice for channel-independent reconstruction, while dual-domain approaches are required for joint multi-coil processing (Souza et al., 2019).
• Fusion Placement: Early, late, or iterative fusion each have empirical trade-offs. For Y-Net, bottleneck fusion sufficed, whereas DBF-Net and KV-Net require fusion after every block or cascade for maximal effect (Xu et al., 2024, Liu et al., 2022).
• Parameter Efficiency: Multi-domain adapters (3D U²-Net) can achieve massive parameter reduction vs. fully replicated models with only a small compromise in accuracy (Huang et al., 2019).
• Task Adaptability: Modular design with task-specific branches or adapters enables efficient extension to new domains with minimal retraining (Huang et al., 2019).
• Branch Supervision: Auxiliary losses on all outputs (e.g., body and boundary maps, intermediate reconstructions) significantly improve convergence and final metrics (Xu et al., 2024).

7. Perspectives and Generalization Potential

The dual-domain/branch U-Net methodology demonstrates broad synthesis and extensibility:

• For any imaging task characterized by separable cues—whether by representation, modality, or abstracted semantics—dual-branch architectures offer a principled, empirically validated route to enhanced performance.
• Feature fusion strategies (auxiliary networks, learnable weights, attention, RL-guided alignment) are application-specific but generalizable.
• This approach also underpins modern universal models for multi-domain learning, allowing a single network to accommodate diverse datasets or tasks with minor architectural overhead (Huang et al., 2019).

A plausible implication is that future segmentation and reconstruction frameworks will increasingly adopt dual-branch principles, combining physically grounded, domain-encoded processing with adaptive feature fusion tailored to application context and dataset diversity.