SEVector: Lightweight Channel Attention

• SEVector is a lightweight channel attention mechanism that recalibrates CNN channel features using dual global pooling and a compact MLP, ensuring minimal overhead.
• It enhances model performance in medical imaging by improving intra-class compactness and inter-class separability, leading to notable accuracy and macro F1 improvements.
• Empirical studies report increased accuracy, stable ROC/AUC metrics, and efficient convergence on CPU-restricted environments, highlighting its practical benefits.

A lightweight channel attention mechanism improves the representational capacity of convolutional neural networks (CNNs) with minimal computational and parameter overhead. The Squeeze-and-Excitation Vector (SEVector) module exemplifies this design paradigm by combining global pooling with a compact multi-layer perceptron, recalibrating channel representations in a resource-efficient manner. Recent studies have validated SEVector’s applicability for medical image analysis in CPU-restricted environments and benchmarked it against alternative lightweight channel-attention strategies.

1. Architectural Overview and Motivation

SEVector is a streamlined variant of the classic Squeeze-and-Excitation (SE) block, designed to minimize parameter growth and computational demands while preserving the core principle of channel-adaptive recalibration. Rather than operating on the $C\times H\times W$ tensor with conventional large fully connected layers, SEVector pools the $C$-channel post-convolutional feature maps $F\in\mathbb{R}^{C\times H\times W}$ into two global descriptors—via Global Average Pooling (GAP) and Global Max Pooling (GMP)—to capture both stable contextual and salient activation information. These are concatenated to form a $2C$-dimensional channel summary vector, providing a dual perspective on channel statistics.

SEVector is specifically motivated by the need for discriminative, robust features in medical imaging under low-compute constraints, facilitating intra-class compactness and inter-class separability without architectural complexity inflation (Xia et al., 15 Aug 2025, Dahes, 6 Dec 2025).

2. Mathematical Formulation and Data Flow

The operation of SEVector consists of several well-defined stages:

1. Dual Pooling Fusion
   • Global average pooling:

   $$v_{\mathrm{avg}} = \mathrm{GAP}(F) = \frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}F_{:,i,j}\in\mathbb{R}^C$$

• Global max pooling:

  $$v_{\mathrm{max}} = \mathrm{GMP}(F) = \max_{i,j}F_{:,i,j}\in\mathbb{R}^C$$

• Concatenation:

  $$v_{\mathrm{fused}} = [v_{\mathrm{avg}}\,;\,v_{\mathrm{max}}]\in\mathbb{R}^{2C}$$

1. Bottleneck Squeeze
   • Reduction to $d = \max\!\left(8, \frac{2C}{r}\right)$ units ($r=16$):

   $$z = W_1\,v_{\mathrm{fused}}\,,\quad W_1\in\mathbb{R}^{d\times 2C}$$

• ReLU nonlinearity:

  $s = \mathrm{ReLU}(z)\in\mathbb{R}^d$

1. Excitation
   • Projection back to $2C$:

   $e = W_2\,s\,,\quad W_2\in\mathbb{R}^{2C\times d}$

• Sigmoid gating:

  $\omega = \sigma(e)\in\mathbb{R}^{2C}$

1. Recalibration
   • Element-wise product:

   $$v_{\mathrm{att}} = v_{\mathrm{fused}} \odot \omega\in\mathbb{R}^{2C}$$

SEVector is typically inserted immediately after the pooled feature fusion and prior to the classification head in architectures such as ConvNeXt-Tiny and InceptionV3 (Xia et al., 15 Aug 2025, Dahes, 6 Dec 2025).

3. Parameter Complexity and Comparative Analysis

The parameter cost of SEVector is governed by its two linear transformations: $$\mathrm{Param}_{\mathrm{SEVector}} = 2(2C\cdot d) = 4Cd$$ For $d=2C/r$ (assuming $2C>8r$), this yields $8C^2/r$ total parameters for the block. However, since SEVector is employed only once atop the feature extraction backbone (as opposed to pervasive block-wise insertion), its fractional contribution to the total model parameter count remains negligible. For example, ConvNeXt-Tiny with $C=96$ and $d=12$ results in approximately $4{,}608$ extra parameters ($<0.1\%$ of a $\sim5$M-parameter model) (Xia et al., 15 Aug 2025).

In multiclass medical image classification, the cost–benefit profile is favorable: integrating SEVector raises accuracy by $6.7$ percentage points over baseline ConvNeXt-Tiny, decreasing validation loss by $18\%$ at $10$ epochs, with convergence confirmed on multi-threaded CPU settings (Xia et al., 15 Aug 2025).

4. Empirical Performance and Impact in Medical Imaging

SEVector has been empirically validated on diverse medical datasets—Alzheimer MRI and consolidated mammography (INbreast, MIAS, DDSM). In all cases, incorporating SEVector after dual pooling fusion led to:

• Increased accuracy and macro F1 (up to $0.985$ and $0.979$ in mammography) (Dahes, 6 Dec 2025)
• Stability in recall, especially for minority or clinically significant classes (e.g., malignant lesions)
• Tighter feature-space clustering (as evidenced by PCA/t-SNE) and improved boundary clarity
• Robust discrimination in terms of ROC/AUC, with AUC for malignant cases sustained at $0.99$

A plausible implication is that SEVector’s explicit channel reweighting acts as a form of feature regularization, suppressing noise and highlighting clinically relevant activations for each class.

Model	Accuracy	Macro F1	Recall_min	AUC (malignant)
Baseline (no SEVector)	0.965	0.958	0.942	0.99
ICNT (GAGM+SEVector)	0.985	0.979	0.966	0.99

5. Implementation, Guidelines, and Architectural Considerations

Key practical aspects for deployment include:

• Pooling fusion: Concatenate both global average and max-pooled descriptors to capture a broad context; this is critical for balancing robustness and salience.
• Module insertion: Apply SEVector after pooling-fusion, before the classifier head, to minimize memory and computation.
• Reduction ratio: Default to $r=16$, which maintains expressivity yet constrains parameter growth. Set the bottleneck $d\ge8$ to avoid collapse in slim models.
• Parameter overhead: For wider spatial backbones (e.g., $C=768$ for ConvNeXt), the increase is $\sim0.3$M parameters.
• Integrated schema: A typical pattern is $${\rm backbone} \rightarrow {\rm GAP+GMP} \rightarrow {\rm Concatenate} \rightarrow {\rm SEVector} \rightarrow {\rm Classifier}$$.
• Extensibility: Architectures can append lightweight spatial attention (e.g., CBAM spatial module) if location modeling is required for small or off-center targets (Dahes, 6 Dec 2025).
• Other modalities: Transfer to modalities (CT, ultrasound) may require adapting the squeeze step if channel semantics or statistical properties differ significantly.

6. Comparison with Alternative Lightweight Channel Attention Mechanisms

SEVector lies within the family of lightweight channel-attention methods:

• SE Block [Hu et al., 2017]: Operates on a single compressed channel summary using a two-layer FC bottleneck; parameter cost increases rapidly with $C^2/r$ for each block.
• ECA Module: Avoids dimensionality reduction, applying 1D convolutions to pooled vectors for local channel interaction. With only $O(k)$ parameters ($k$ small), ECA achieves similar or superior accuracy (e.g., $+2.28\%$ Top-1 on ImageNet over ResNet-50) at virtually no cost (Wang et al., 2019).
• LCT Block: Utilizes group normalization and per-channel affine transforms, reducing parameter count to $O(C)$, achieving ImageNet and COCO gains comparable to or exceeding SE/SEVector with almost zero overhead (Ruan et al., 2019).
• CRA Module: Incorporates spatial information by using depthwise convolutions on small pooled grids; outperforms SE/SEVector in some settings by using fewer parameters and recovering spatial structure (Shen et al., 2020).
• Multi-axis SEVector: Designs such as tfwSE and channel-frequency pooling, explored in audio SED, demonstrate that “SEVector” patterns (global pool $\rightarrow$ bottleneck FC $\rightarrow$ sigmoid $\rightarrow$ rescale) can be deployed along any axis with $\ll5\%$ parameter addition, recapturing large-model performance at a fraction of the cost (Nam et al., 2023).

7. Limitations and Prospective Directions

Limitations of SEVector include:

• Lack of spatial attention: Fine-grained or localized lesions may require additional spatial modules.
• Residual class imbalance: Channel reweighting alone cannot address systemic class imbalance.
• Reduced benefits on shallow networks: Narrow or shallow backbones see smaller performance improvements.

Future directions entail integrating spatially-adaptive attention at minimal cost, further compressing the bottleneck (perhaps with advanced quantization or pruning), and adaptive pooling strategies tailored to medical or domain-specific statistics. A plausible implication is that combining SEVector with niche-aware data augmentation and spatial mechanisms could further enhance compactness and separability for challenging classification scenarios.

• "An Efficient Medical Image Classification Method Based on a Lightweight Improved ConvNeXt-Tiny Architecture" (Xia et al., 15 Aug 2025)
• "Proof of Concept for Mammography Classification with Enhanced Compactness and Separability Modules" (Dahes, 6 Dec 2025)
• "Linear Context Transform Block" (Ruan et al., 2019)
• "ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks" (Wang et al., 2019)
• "Convolutional Neural Network optimization via Channel Reassessment Attention module" (Shen et al., 2020)
• "Frequency & Channel Attention for Computationally Efficient Sound Event Detection" (Nam et al., 2023)