DBIF-AUNet: Dual-Branch Fusion for CT Segmentation

• The paper introduces a dual-branch interactive fusion model that decouples and adaptively fuses global, local, and frequency features to overcome semantic gaps in CT segmentation.
• Densely nested skip connections and the DDFD module refine feature extraction, improving boundary delineation and robustness against low contrast and blurred edges.
• A nested deep supervision with a hybrid loss achieves an IoU of 80.1% and a Dice score of 89.0%, outperforming U-Net++ and Swin-UNet in pleural effusion segmentation.

Searching arXiv for DBIF-AUNet and closely related dual-branch interactive fusion attention papers to ground the article with current references. arXiv search query: "DBIF-AUNet" Dual-Branch Interactive Fusion Attention Model (DBIF-AUNet) is a semantic segmentation architecture for pleural effusion CT images that addresses low contrast, blurred boundaries, variable morphology, and the semantic gaps induced by direct feature concatenation. It builds on a densely nested skip-connection network, refines a Dual-Domain Feature Disentanglement module (DDFD), introduces a Branch Interaction Attention Fusion module (BIAF), and adds a nested deep supervision mechanism with hierarchical adaptive hybrid loss. On 1,622 pleural effusion CT images from Southwest Hospital, it achieved an IoU of 80.1% and a Dice score of 89.0%, outperforming U-Net++ and Swin-UNet by 5.7%/2.7% and 2.2%/1.5% respectively (Tang et al., 8 Aug 2025).

1. Problem setting and design objective

Pleural effusion semantic segmentation is framed as a clinically relevant CT segmentation task in which accurate delineation of lesion areas can enhance the accuracy and timeliness of diagnosis and treatment. The task is difficult because effusion and surrounding tissues may have similar gray levels, edges are often blurred, and lesion morphology is variable. The model is explicitly motivated by the claim that existing methods struggle with diverse image variations and complex edges, primarily because direct feature concatenation causes semantic gaps (Tang et al., 8 Aug 2025).

DBIF-AUNet targets these failure modes by replacing straightforward skip-feature aggregation with structured disentanglement and adaptive fusion. Its design premise is that multiscale features should not merely be concatenated across encoder and decoder stages, but first be re-aligned and re-weighted so that global structure, local boundary detail, and frequency-sensitive texture cues can contribute in a coordinated way. This makes the architecture part of a broader shift away from static feature fusion toward interaction-aware fusion.

2. Densely nested architecture and network composition

DBIF-AUNet builds on the U-Net++ family through densely nested skip connections. These connect encoder and decoder layers across multiple semantic scales and are intended to reduce the semantic gap between features from different depths, improving multiscale feature blending (Tang et al., 8 Aug 2025).

The architecture is organized into five main components:

Component	Role	Main function
Encoder	Feature extraction	Produces multiscale representations
Decoder	Reconstruction	Restores segmentation resolution
DDFD	Feature disentanglement	Separates global, local, and channel cues
BIAF	Interactive fusion	Dynamically weights and fuses branches
Nested deep supervision branches	Optimization	Supplies multi-level training signals

In this configuration, features from different depths are not just concatenated. They are processed and re-aligned through DDFD and BIAF for robustness under fuzzy boundaries and similar tissue intensities. At the core of the model is the fusion of features across multiple branches—specifically, global (low-frequency/structure), local (detail/edge), and channel (frequency/textural)—after disentanglement by DDFD and adaptive fusing by BIAF. This contrasts with static concatenation or addition in previous U-Net variants (Tang et al., 8 Aug 2025).

A common simplification is to treat DBIF-AUNet as merely a nested U-Net with attention. That description is incomplete. The distinctive architectural claim is the combination of dense nesting with orthogonal feature decoupling and interactive weighting, rather than attention appended to an otherwise conventional encoder–decoder.

3. Dual-Domain Feature Disentanglement (DDFD)

The DDFD module is introduced to address the limitation that simple spatial feature aggregation often fails in pleural effusion CT because of low contrast, blurry boundaries, and heterogeneous object morphology. Its stated solution is to orthogonally decouple global, local, and channel-wise features, integrating both spatial and frequency domain information (Tang et al., 8 Aug 2025).

The module consists of three functionally decoupled submodules operating on aggregated features from nested skip connections.

The global branch captures global or low-frequency context. It applies global average pooling to extract overall image morphology, then uses a Discrete Cosine Transform (DCT) and a learned frequency-domain channel attention mechanism to emphasize informative frequency-band context.

The local branch enhances edges, high-frequency detail, and boundary response. It uses spatial strip pooling to capture elongated structures, then applies a 2D Discrete Wavelet Transform (DWT) to split features into low-frequency and high-frequency components, denoted $X_{LL}$ and $X_{HH}$. Gabor filtering is then used for frequency-selective edge enhancement:

$$\mathbf{G}_{\theta,f}(i,j) = \exp\left( -\frac{i'^2 + \gamma^2 j'^2}{2\sigma^2} \right)\cos(2\pi f i' + \phi)$$

The channel branch enriches channel and frequency texture information through a 2D DCT decomposition followed by inverse reconstruction. Its frequency-domain representation is written as

$$F(u,v) = \alpha(u)\alpha(v) \sum_{i=0}^{H-1}\sum_{j=0}^{W-1} X_{c,i,j} \cos\left[\frac{\pi(2i+1)u}{2H}\right] \cos\left[\frac{\pi(2j+1)v}{2W}\right]$$

The stated effect of this organization is orthogonal decoupling and multiscale complementarity: global structure, sharp boundaries, and channel texture are modeled separately so that redundancy and confusion are reduced before fusion. This design is central to the model’s interpretation of “interactive fusion,” because interaction occurs after representational roles have first been disentangled rather than mixed prematurely.

4. Branch Interaction Attention Fusion (BIAF)

After DDFD, DBIF-AUNet uses the Branch Interaction Attention Fusion module to dynamically combine the enhanced branch outputs. The paper argues that traditional fusion by simple sum or concatenation is suboptimal because the relevance of global, local, and channel/frequency information depends on lesion shape and local context (Tang et al., 8 Aug 2025).

BIAF is described as a three-stage process.

First, parallel enhancement is applied to the three branches. The global branch uses frequency-domain self-attention for long-range feature correlation with cross-band complementarity. The local branch uses dynamic convolution and frequency band normalization to handle deformation and distribution differences. The channel branch uses deformable convolution together with learnable soft channel thresholds to suppress irrelevant frequency noise.

Second, dynamic fusion computes branch-specific attention maps and concatenates branch outputs, followed by fully connected layers to produce fused features:

$$\mathbf{Y} = \sigma( W_{f_2}^T \cdot ReLU( W_{f_1}^T \cdot GAP(Z) ) ) + \mathrm{Concat}(A_g * X_1, A_l * X_2, A_c * X_3)$$

Third, interactive weighting normalizes branch contributions with Softmax:

$$B_i = \frac{\exp(W_{b_i}^T \cdot Y)}{\sum_{j=1}^3 \exp(W_{b_j}^T \cdot Y)} \quad \text{for}~i=1,2,3$$

and the final output is

$O = \sum_{i=1}^3 B_i \odot Y_i$

The significance of BIAF lies in its replacement of fixed fusion rules with data-dependent weighting. The fused representation is explicitly defined as an adaptive combination of global, local, and textural information, rather than a uniform merge. In the paper’s framing, this is the mechanism that improves robustness under blurred edges and variable morphologies.

5. Nested deep supervision and hierarchical adaptive hybrid loss

DBIF-AUNet augments feature fusion with a nested deep supervision mechanism. Beyond the standard U-Net++ practice of attaching losses to decoder stages, it inserts supervision heads after each BIAF output and at each upsampling path. Each supervision loss is adaptively weighted according to spatial resolution and layer depth, with Softmax normalization for stability (Tang et al., 8 Aug 2025).

At each supervision head, the model uses a hybrid loss

$$\mathcal{L}_\text{hybrid}^{(k)} = \lambda_\text{Dice}\,\mathcal{L}_\text{Dice}^{(k)} + \lambda_\text{Focal}\,\mathcal{L}_\text{Focal}^{(k)} + \lambda_\text{BCE}\,\mathcal{L}_\text{BCE}^{(k)}$$

with

$$\lambda_{Dice}=0.4,\quad \lambda_{Focal}=0.3,\quad \lambda_{BCE}=0.3$$

The Dice term is

$$\mathcal{L}_\text{Dice}^{(k)} = 1 - \frac{2 \sum (Y_\text{true} \odot Y_\text{pred}^{(k)}) + \epsilon}{\sum Y_\text{true} + \sum Y_\text{pred}^{(k)} + \epsilon}$$

The Focal term is

$X_{HH}$0

The binary cross-entropy term is

$X_{HH}$1

The total loss aggregates decoder-stage and BIAF-stage losses with depth- and area-adaptive weights:

$X_{HH}$2

The paper attributes several functions to this design: Dice and Focal address class imbalance and ambiguous borders, Focal emphasizes hard-to-segment pixels, BCE stabilizes optimization, and weighted multilevel supervision improves performance on small and fuzzy effusion regions. One practical implication is that the model’s gains are not presented as arising from architectural depth alone; the optimization scheme is treated as a co-equal component of the method.

6. Empirical results, ablation evidence, and relation to adjacent literature

DBIF-AUNet was validated on 1,622 expert-annotated pleural effusion CT images with high consistency and diverse cases. The reported comparison is as follows (Tang et al., 8 Aug 2025):

Model	IoU (%)	Dice (%)
U-Net	71.1	83.1
UNet++	74.4	85.3
ANU-Net	75.1	85.8
Swin-UNet	77.9	87.6
DBIF-AUNet	80.1	89.0

Other metrics—Accuracy, Precision, Recall, and Specificity—are also reported as highest for DBIF-AUNet. Relative improvements are explicitly given as +2.2% IoU and +1.5% Dice over Swin-UNet, and +5.7% IoU and +2.7% Dice over UNet++.

The ablation study is particularly important for interpreting the model. Removing DDFD+BIAF reduces IoU from 80.1% to 74.6%, while removing deep supervision reduces IoU to 79.5%. This supports a specific reading of the method: the performance gain is not reducible to the nested backbone alone. The key gain is attributed to dual-branch fusion and enhanced supervision. Qualitatively, the paper states that boundary delineation is sharper, more resistant to noise and blurred regions, and better able to handle small, diffused, large, or encapsulated effusions than the baselines.

Within adjacent arXiv literature, dual-branch interactive fusion is instantiated in several distinct ways. The Dual Interaction Network introduces a dual interactive fusion module using bidirectional cross-attention between original and enhanced images plus global spatial attention for medical segmentation (Noh et al., 7 Sep 2025). DRIFA-Net separates multi-branch fusion attention for unimodal enhancement from multimodal information fusion attention for shared representations across dermoscopy, pap smear, MRI, and CT-scan (Dhar et al., 2024). BiDGANet uses a bilateral architecture with Dual-Guided Attention to fuse high-resolution spatial and low-resolution semantic features under real-time constraints (Liao et al., 2023). DB-KAUNet adopts a heterogeneous CNN/Transformer encoder with Cross-Branch Channel Interaction and spatial enhancement modules for retinal vessel segmentation (Xu et al., 1 Dec 2025). This suggests that DBIF-AUNet belongs to a broader family of architectures that replace static concatenation with learned interaction, but its specific contribution is the coupling of densely nested skip connections with orthogonal feature disentanglement and adaptive global–local–frequency fusion for pleural effusion CT segmentation.

A recurring misconception in discussions of such models is that “dual-branch” necessarily denotes only two homogeneous streams. In DBIF-AUNet, the nomenclature coexists with a finer internal decomposition into global, local, and channel/frequency branches inside DDFD and BIAF. The model is therefore more accurately understood as a dual-branch interactive fusion framework whose internal attention machinery explicitly separates and recombines multiple representational domains.