Channel Concatenate & Element Add (CA) Fusion

• Channel Concatenate and Element Add (CA) is a lightweight decoder fusion block that refines segmentation boundaries by integrating low-level spatial detail with upsampled semantic features.
• It employs dual 1×1 convolutions to project and split low-level features, then fuses one branch via concatenation with a convolution and the other via element-wise addition.
• Ablation studies on the Kvasir-SEG dataset indicate that CA modestly improves Dice and IoU metrics, confirming its role in enhancing boundary reconstruction in polyp segmentation.

Channel Concatenate and Element Add (CA) is the lightweight decoder-side fusion block introduced in CL-Polyp, a contrastive learning-enhanced network for polyp segmentation built on a DeepLabV3+ style encoder-decoder architecture with a ResNet50 encoder (Li et al., 9 Jul 2025). Within that model, CA operates in the segmentation branch after high-level features have been extracted by the encoder and after the decoder has begun to upsample them. Its purpose is to improve the fusion of low-level spatial detail from the encoder with upsampled semantic features from the decoder, thereby improving boundary reconstruction for polyps whose edges are blurry, thin, fuzzy, or low-contrast.

1. Architectural role in CL-Polyp

In CL-Polyp, CA is not the main semantic feature extractor. The model remains primarily encoder-focused, while the decoder is modified to recover boundary detail better. The decoder introduces two custom modules: the Modified Atrous Spatial Pyramid Pooling (MASPP) module for stronger multi-scale feature aggregation at deeper layers, and the Channel Concatenate and Element Add module for better fusion of low-level and upsampled features during decoding (Li et al., 9 Jul 2025).

This placement gives CA a specific functional role. It acts as a feature refinement and boundary reconstruction module in the decoder, rather than as a replacement for the encoder or for the model’s contrastive learning-enhanced representation learning. The paper presents it as a way to make the skip connection more expressive than a plain one-to-one transfer. A plausible implication is that CA is intended to address the known mismatch between semantically strong but spatially coarse decoder features and the high-resolution spatial cues preserved in earlier encoder stages.

2. Fusion rule and tensor operations

The paper describes CA as a two-step replacement for the simpler DeepLabV3+ skip fusion. A low-level encoder feature map is first split or projected into two parts using two $1 \times 1$ convolutions, and one projected branch is then combined with the upsampled decoder feature by channel concatenation, followed by convolution, after which the other projected branch is injected by element-wise addition (Li et al., 9 Jul 2025).

The low-level input is described as

$f_2 \in \mathbb{R}^{C \times H \times W}.$

Two projected tensors are produced:

$$A_i = \mathrm{Conv}_{1\times1}^{(A)}(f_2), \qquad B_i = \mathrm{Conv}_{1\times1}^{(B)}(f_2).$$

The decoder feature $z_i$ is upsampled to the same spatial resolution as $f_2$, and the core CA fusion equation is

$z_i = A_i + \mathrm{Conv}(\mathrm{cat}(B_i, z_i)).$

The mechanism implies the following computation sequence. First, $f_2$ is projected twice into channel-compatible tensors. Second, $B_i$ and the upsampled $z_i$ are concatenated along the channel axis. Third, a convolution is applied to the concatenated tensor to align and refine the fusion. Finally, the result is element-wise added to $A_i$. The paper does not provide a longer formal derivation, but it makes clear that concatenation is performed along the channel dimension, convolution follows concatenation, and addition completes the fusion (Li et al., 9 Jul 2025).

This structure is explicitly hybrid. It combines concatenation-based fusion, which preserves and mixes information from the decoder and one low-level projection, with an additive low-level pathway that acts as a residual-style refinement.

3. Boundary reconstruction and segmentation rationale

The motivation for CA is tied to the visual properties of polyp segmentation. The paper argues that polyp boundaries are often fuzzy, low-contrast, and similar in appearance to surrounding tissue. In such cases, decoder features alone may be semantically strong but spatially coarse, while low-level encoder features preserve edges and local texture (Li et al., 9 Jul 2025).

CA is presented as beneficial in two related ways. First, it makes better use of low-level detail. Because the low-level feature map is projected into two branches, the module does not force all boundary-relevant information through a single direct skip path. Second, it promotes stronger interaction between semantic and spatial cues. The concatenation branch mixes decoder semantics with a projected low-level component, while the addition branch provides a direct additive refinement.

The paper links this design to the morphology and appearance of polyps, which often exhibit irregular shape, small size, unclear borders, and background similarity. In practice, CA is intended to make the decoder’s upsampling step less lossy and more boundary-sensitive. This suggests that its contribution is most pronounced at object contours rather than in coarse region localization.

4. Relation to standard skip fusion and simpler alternatives

The paper explicitly contrasts CA with the standard DeepLabV3+ skip connection. In the baseline decoder, a low-level feature map from an earlier encoder stage is directly fused with the upsampled decoder output in a relatively straightforward way. CA replaces that simpler mechanism with a structured split-project-fuse-add pattern (Li et al., 9 Jul 2025).

Compared with standard skip fusion, CA uses two projected parts of the low-level feature rather than a single low-level pathway. Compared with concatenation-only fusion, it does not rely solely on later convolution to learn the entire fusion; it also adds a separate low-level projection $f_2 \in \mathbb{R}^{C \times H \times W}.$0 directly to the fused output. Compared with addition-only fusion, it retains the flexibility of concatenation rather than requiring tightly aligned channel-wise fusion as the sole interaction mechanism.

The paper therefore characterizes CA as a combined concat-and-add fusion strategy intended to be more expressive than either operation alone. Its significance within CL-Polyp is not that it replaces semantic encoding, but that it refines decoder-side feature merging in a way specifically targeted at boundary recovery.

5. Ablation findings and empirical contribution

The paper includes an ablation study on the Kvasir-SEG dataset that isolates the incremental effect of MASPP, CA, and the contrastive learning branch (Li et al., 9 Jul 2025).

Method	Dice	IoU
Baseline	0.889	0.827
Baseline + MASPP	0.899	0.838
Baseline + MASPP + CA	0.900	0.842
Baseline + MASPP + CA + CL Branch	0.918	0.864

From this study, the paper states that adding the CA module on top of MASPP yields a 0.1% improvement in Dice and a 0.4% improvement in IoU. The paper describes this as a modest but real gain, consistent with CA’s role as a refinement module rather than a major semantic extractor.

Within the full CL-Polyp system, the model is reported to improve the IoU metric by 0.011 on Kvasir-SEG and 0.020 on CVC-ClinicDB relative to the compared state-of-the-art methods. Those gains validate the overall architecture rather than CA alone, but the ablation indicates that improved boundary fusion contributes measurably to the final performance (Li et al., 9 Jul 2025).

6. Terminological scope and distinctions from similarly abbreviated mechanisms

A recurrent source of ambiguity is that “CA” and related abbreviations are used differently across segmentation and image reconstruction literature. In CL-Polyp, CA denotes a decoder-side fusion block whose defining operation is

$f_2 \in \mathbb{R}^{C \times H \times W}.$1

with two $f_2 \in \mathbb{R}^{C \times H \times W}.$2 low-level projections, channel concatenation, convolution, and element-wise addition (Li et al., 9 Jul 2025).

This is distinct from Channelized Axial Attention (CAA), which is not a channel-concatenate-plus-add block. CAA reformulates axial spatial attention by inserting channel relation inside the spatial attention computation itself. Its core mechanism uses attention decomposition and channel-wise gating by element-wise multiplication, not concatenation-plus-addition as the main operation (Huang et al., 2021).

It is also distinct from the “CA” mechanism in PyNET-CA. There, the core architectural change is the MultiConv channel attention block, where outputs of multiple convolution branches are concatenated and then reweighted by channel attention. The explicitly defined element-wise operation is multiplication of attention weights with feature channels, and the paper does not give a separate explicit equation for element-wise additive gating inside the CA block (Kim et al., 2021).

These distinctions matter because the CL-Polyp module is best understood as a decoder fusion operator for low-level and upsampled features, not as a generic label for channel attention, axial attention, or multibranch recalibration. Its specific contribution lies in boundary-oriented skip refinement within a polyp segmentation decoder.