Adaptive SE Res U-Net

• ASE_Res_UNet is a U-Net–derived segmentation architecture that unifies residual encoding with adaptive squeeze-and-excitation decoding for robust biomedical segmentation under noisy conditions.
• It strategically uses residual blocks in the encoder and noise-aware SE modules in the decoder to preserve spatial detail and enhance weak signal recovery.
• Empirical evaluations demonstrate improved Dice and sensitivity metrics on challenging datasets, highlighting its effectiveness in segmenting curvilinear structures.

Searching arXiv for the cited ASE_Res_UNet paper and closely related SE/Residual U-Net work to ground the article in current papers. arXiv search query: "Adaptive Squeeze-and-Excitation Residual U-Net" OR "Squeeze-and-Excitation U-Net" OR "Residual U-Net". Adaptive Squeeze-and-Excitation Residual U-Net (ASE_Res_UNet) denotes a U-Net-derived segmentation architecture that combines residual learning with squeeze-and-excitation-based feature recalibration. In the formulation explicitly introduced under this name, ASE_Res_UNet was developed for segmenting curvilinear structures in noisy and low-contrast biomedical images, notably fluorescent microtubules, retinal vessels, and corneal nerves. Its defining configuration places residual blocks in the encoder and a noise-aware adaptive squeeze-and-excitation module in every decoder block, thereby coupling stable hierarchical feature extraction with input-conditioned channel reweighting during reconstruction (Laydi et al., 10 Jul 2025). As a design family, ASE_Res_UNet sits at the intersection of residual U-Net backbones (Zhang et al., 2017), SE-augmented U-Nets (Rundo et al., 2019), and later segmentation architectures that combine residual blocks, SE attention, and multi-scale context modules (Wang et al., 2021).

1. Origins and architectural lineage

ASE_Res_UNet emerged from the convergence of two lines of work: residual U-Net architectures and squeeze-and-excitation augmentation. Deep Residual U-Net established that U-Net-style encoder-decoder segmentation can be strengthened by replacing plain convolutional units with residual units, while retaining long encoder-decoder skip connections; in that formulation, the network had 15 convolutional layers and 7.8M parameters versus 30.6M for the original U-Net, yet achieved a relaxed precision-recall break-even point of 0.9187 on the Massachusetts Roads Dataset, compared with 0.9053 for U-Net (Zhang et al., 2017). This established the residual U-Net as a compact yet expressive baseline.

A second thread inserted SE modules into U-Net. USE-Net incorporated SE blocks after every encoder block or after every encoder-decoder block for prostate zonal segmentation, and showed that adaptive feature recalibration was especially effective when training on the union of multiple institutional MRI datasets (Rundo et al., 2019). SAR-U-Net then combined residual blocks, SE blocks in the encoder, and ASPP at the bottleneck and output for liver CT segmentation, reporting mean Dice values of 95.71 on LiTS17 and 97.31 on SLiver07 (Wang et al., 2021). Outside segmentation, RSEN demonstrated that a lightweight encoder-decoder built from residual squeeze-and-excitation blocks can provide efficient one-stage restoration, reinforcing the generality of residual-plus-SE design (Fu et al., 2020).

This lineage suggests that ASE_Res_UNet is best understood not as an isolated invention, but as a specific synthesis: residual U-Net topology for optimization and feature propagation, SE-style channel recalibration for adaptive emphasis, and task-driven placement of attention where reconstruction is most vulnerable to noise and class imbalance.

2. Canonical architecture

In its explicit 2025 formulation, ASE_Res_UNet comprises four encoder blocks, four decoder blocks, and a bottleneck (Laydi et al., 10 Jul 2025). The base convolutional unit is a $3\times 3$ convolution followed by Batch Normalization and ReLU, and the initial number of filters is 32, doubling at each downsampling stage. Standard U-Net skip connections concatenate encoder feature maps with upsampled decoder features at corresponding scales. Upsampling is performed with transposed convolutions.

The distinguishing asymmetry is that residual blocks are used in the encoder only, whereas adaptive squeeze-and-excitation modules are used in every decoder block. The encoder design is motivated by the need to preserve gradient flow and subtle spatial detail during repeated downsampling under low signal-to-noise conditions. The decoder design is motivated by the need to adaptively reconstruct fine, low-intensity curvilinear structures from noisy intermediate representations.

The encoder residual block is reported as

$$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$

where $x$ is the block input and $W_1, W_2$ are convolutional kernels (Laydi et al., 10 Jul 2025). Residual blocks are intentionally omitted from the decoder in order to keep the parameter count low.

Component	Reported role	Reported design
Encoder	Hierarchical feature extraction	MaxPooling, two $3\times 3$ Conv-BN-ReLU layers, residual block
Decoder	Detail reconstruction	ConvTranspose upsampling, skip concatenation, two $3\times 3$ Conv-BN-ReLU layers, ASE
Attention placement	Adaptive feature modulation	ASE in every decoder block

This encoder-decoder split is notable because many SE-U-Net variants place SE in the encoder or in both halves. ASE_Res_UNet instead assigns residual stabilization to the downsampling path and adaptive channel recalibration to the upsampling path. This suggests a deliberate division of labor: residual blocks protect representation quality during compression, while adaptive SE governs selective recovery during expansion.

3. Adaptive squeeze-and-excitation mechanism

The ASE module in ASE_Res_UNet is a modified SE block that incorporates a noise estimate into the excitation pathway (Laydi et al., 10 Jul 2025). Standard SE computes channel weights from global average pooled features using a fixed reduction ratio. In ASE_Res_UNet, the reduction ratio becomes data-dependent through a learned noise estimate.

Given an input feature map $X \in \mathbb{R}^{B \times C \times H \times W}$, the module first estimates a noise map:

$N = \text{Sigmoid}(\text{Conv2d}(X))$

and then averages it over batch and spatial dimensions:

$$\bar{N} = \frac{1}{B \times H \times W} \sum_{b=1}^{B} \sum_{i=1}^{H} \sum_{j=1}^{W} N_{b,i,j}$$

The adaptive reduction ratio is then defined as

$$r_\text{adaptive} = \max(1,\lfloor 16 .(1+ \bar{N}) \rfloor)$$

The squeeze stage remains the standard channel-wise global average pooling:

$$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$0

The excitation stage uses two fully connected layers with ReLU and sigmoid:

$$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$1

Finally, the feature map is rescaled channel-wise:

$$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$2

The important distinction from conventional SE is that $$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$3 is not fixed. In standard SE blocks used in USE-Net, the reduction ratio was fixed at $$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$4 (Rundo et al., 2019), and in RSEN the squeezed channel dimension was fixed to 6 (Fu et al., 2020). ASE_Res_UNet instead conditions the internal bottleneck on the estimated noise level. This makes the decoder’s channel attention explicitly noise-aware rather than merely input-aware.

The paper attributes the model’s gains in low-intensity and high-noise regions largely to this adaptive SE design (Laydi et al., 10 Jul 2025). A plausible interpretation is that ASE converts channel attention from a purely representational prior into a reconstruction control mechanism that changes capacity with the corruption level of the current feature tensor.

4. Data regime, optimization, and evaluation

ASE_Res_UNet was introduced and evaluated primarily on two synthetic microtubule datasets, MicSim_FluoMT simple and MicSim_FluoMT complex, each containing 1192 images with train/validation/test splits of 953/119/120 (Laydi et al., 10 Jul 2025). The simple dataset has uniform fluorescence and an SNR of approximately 1.7 dB, whereas the complex dataset has fluorescence intensity that gradually decreases toward the cell periphery and an SNR of approximately 1.2 dB. This complex variant is the principal stress test for the architecture because filament extremities become faint and approach the background distribution.

The same study also evaluates on MicReal_FluoMT, a real stained-microtubule dataset with 49 image-mask pairs split into 19 training, 10 validation, and 10 test images; on DRIVE, with a 20/10/10 split; and on CORN-1 using 5-fold cross-validation with training on 80% of data in each fold (Laydi et al., 10 Jul 2025). The paper emphasizes that the main difficulties are high noise, dense filament networks, low contrast, inhomogeneous intensity, and extreme foreground-background imbalance.

Training uses Adam with $$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$5, $$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$6, an initial learning rate of $$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$7, and batch size 2 (Laydi et al., 10 Jul 2025). Training stops when validation loss stabilizes over 50 epochs. The principal loss is Weighted Cross Entropy:

$$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$8

with empirically fixed class weights $$y= \text{ReLU}(\text{BN}(W_2 * \text{ReLU}(\text{BN}(W_1 * x))) + x)$$9 for foreground and $x$0 for background. The study also evaluates BCE, Dice Loss, Focal Loss, and a Hausdorff-distance-based loss, but selects WCE for the main experiments (Laydi et al., 10 Jul 2025).

Evaluation emphasizes imbalance-robust metrics: Dice, IoU, Sensitivity, Precision, MCC, and PR AUC. The authors explicitly avoid relying on accuracy, specificity, and ROC AUC for model comparison in this setting because background pixels dominate the image and can make weak segmenters appear deceptively strong (Laydi et al., 10 Jul 2025).

5. Empirical behavior and ablation structure

On the easier MicSim_FluoMT simple dataset, all variants perform similarly. The reported Dice values are $x$1 for U-Net and $x$2 for ASE_Res_UNet, indicating that the added mechanisms are not decisive when microtubules are clearly visible (Laydi et al., 10 Jul 2025). The architecture’s distinct behavior appears on the complex dataset, where low-intensity extremities and noise dominate the failure modes.

Model	Dice	Sensitivity
U-Net	$x$3	$x$4
ASE_UNet	$x$5	$x$6
ASE_Res_UNet	$x$7	$x$8

On this complex dataset, ASE_Res_UNet also reports IoU $x$9, Precision $W_1, W_2$0, MCC $W_1, W_2$1, and PR AUC $W_1, W_2$2 (Laydi et al., 10 Jul 2025). The confusion matrices further show a reduction in false negatives, from 792 for U-Net to 548 for ASE_Res_UNet, with true positives increasing from 1281 to 1525. In spatially stratified evaluation, the peripheral region is the most informative: ASE_Res_UNet reports Dice $W_1, W_2$3, Sensitivity $W_1, W_2$4, MCC $W_1, W_2$5, and PR AUC $W_1, W_2$6, outperforming the other variants in the weakest-signal zone (Laydi et al., 10 Jul 2025).

Against external baselines on the same complex dataset, ASE_Res_UNet is reported as the best overall model among CAR-UNet, AG_Res_UNet, SE_Res_UNet, Pix2pix, and TransUNet, while using approximately 7.96M parameters, compared with 15.83M for CAR-UNet and 32.6M for TransUNet (Laydi et al., 10 Jul 2025). The same architecture generalizes to real and cross-domain settings: on MicReal_FluoMT it reports IoU $W_1, W_2$7, Sensitivity $W_1, W_2$8, and Precision $W_1, W_2$9; on DRIVE it reports Dice $3\times 3$0 and PR AUC $3\times 3$1; on CORN-1 it reports Dice $3\times 3$2 and PR AUC $3\times 3$3 (Laydi et al., 10 Jul 2025).

These ablations support a specific interpretation. Residual encoding alone improves optimization and representation stability, ASE alone raises recall on weak structures, and the combination yields the strongest reduction in missed filament pixels. The architecture is therefore not merely additive in the sense of “residual plus attention,” but complementary in failure mode: residual blocks counter degradation in deep encoding, whereas adaptive SE compensates for noise and intensity heterogeneity during decoding.

6. Variants and neighboring formulations

The broader literature shows that “ASE_Res_UNet” names a design space rather than a single immutable module. USE-Net adds classic SE blocks after every encoder block or after every encoder-decoder block, showing that encoder-decoder-wide SE can be especially effective when trained on the union of heterogeneous institutional datasets (Rundo et al., 2019). SAR-U-Net places SE in the encoder of a residual U-Net and supplements it with ASPP at the bottleneck and output, reporting mean Dice values of 95.71 on LiTS17 and 97.31 on SLiver07 (Wang et al., 2021). A different branch replaces fixed affine normalization parameters with SE-generated $3\times 3$4 and $3\times 3$5, yielding Squeeze-and-Excitation Normalization in a residual 3D U-Net for PET/CT tumor delineation (Iantsen et al., 2021).

Other work generalizes SE beyond channel-only recalibration. Spatially-Adaptive Squeeze-Excitation Networks define spatially varying attention tensors rather than channel-wise scalars, framing SASE as a convolutional alternative to multi-head self-attention (Shen et al., 2021). U-RWKV introduces a stage-adaptive squeeze-and-excitation module whose internal structure changes between shallow and deep stages, and reports that removing SASE lowers average IoU from 77.62 to 75.29 in the reported ablation (Ye et al., 15 Jul 2025). Search-based work decomposes attention into channel squeeze, channel excitation, spatial squeeze, and spatial excitation operations, then uses differentiable NAS to discover effective SE-like blocks (Wang et al., 2024).

Variant	Adaptive mechanism	Reported placement
USE-Net	Classic channel SE	After encoder blocks, or after encoder-decoder blocks
SAR-U-Net	Classic channel SE + ASPP	Encoder only
SE Norm U-Net	SE-conditioned normalization	After conv+ReLU in encoder and decoder
SASE / stage-adaptive SE	Spatially or stage-varying excitation	Residual blocks or stage modules

This suggests that ASE_Res_UNet is best treated as a family of residual U-Nets whose adaptive component may act through fixed channel gating, noise-conditioned excitation, normalization parameters, spatially varying tensors, or stage-dependent internal structure. The 2025 ASE_Res_UNet corresponds to the noise-aware decoder-gating member of that family (Laydi et al., 10 Jul 2025).

7. Limitations and open directions

The explicit ASE_Res_UNet paper reports several limitations. Even the full model struggles with extremely faint filament extremities that are visually indistinguishable from background, and quantitative assessment on real microtubule data is constrained by noisy semi-supervised labels around centrosomes, embryo peripheries, and non-specific staining (Laydi et al., 10 Jul 2025). Zero-shot transfer to live GFP::TBB-2 imagery reveals domain-shift effects: a model trained on MicReal_FluoMT struggles with cytoplasmic background fluorescence, while a model trained on MicSim_FluoMT complex better handles noise but still exhibits inaccuracies near the centrosome and some hallucinated structures (Laydi et al., 10 Jul 2025).

Related work indicates additional constraints. In USE-Net, the more expressive Enc-Dec SE configuration becomes clearly superior mainly when training on the union of all datasets, which implies that heavy SE deployment benefits from diverse training domains and can otherwise overfit (Rundo et al., 2019). In SAR-U-Net, the authors explicitly identify the 2D nature of the architecture as a limitation because it misses z-axis context in volumetric CT and propose 3D extensions as future work (Wang et al., 2021). In RSEN, SE slightly harmed performance on Rain100H and was disabled for that setting, indicating that channel recalibration is not uniformly beneficial across all forms of structured corruption (Fu et al., 2020).

Taken together, these results imply that the main open problem is not whether residual U-Nets benefit from adaptive excitation, but how adaptation should be parameterized for a given regime. Noise-aware decoder SE, encoder-decoder-wide SE, SE-conditioned normalization, stage-adaptive SE, and spatially adaptive SE are all plausible directions. A plausible next step is therefore a volumetric or multimodal ASE_Res_UNet that combines residual encoding, decoder-side adaptive excitation, and either stage-dependent or spatially adaptive recalibration, while being trained on heterogeneous multi-institutional data so that the adaptive mechanism sees the variability it is expected to normalize.