Hybrid Swin Attention Network

• Hybrid Swin Attention Network (HSANet) is a family of architectures that combine Swin-style hierarchical attention with complementary modules to address limitations like weak fine-grained locality.
• It has been successfully applied in tasks such as low-dose PET/CT denoising, mammography classification, and 3D brain tumor segmentation, demonstrating competitive performance metrics.
• HSANet architectures integrate innovative attention mechanisms, fusion strategies, and specialized training paradigms to optimize global context and local detail extraction for diverse applications.

Searching arXiv for the cited HSANet-related papers and adjacent work to ground the article. Hybrid Swin Attention Network (HSANet) denotes a family of hybrid architectures built around Swin-style hierarchical attention and a complementary mechanism for local detail preservation, feature aggregation, temporal modeling, or efficient decoding. In the literature, the term appears in two distinct ways. First, it is used explicitly as the model name in low-dose PET/CT denoising, where HSANet combines Swin blocks, residual convolutional blocks, Efficient Global Attention modules, and a hybrid upsampling module (Liu et al., 8 Sep 2025). Second, it functions as a conceptual label for architectures whose core design follows the same principle even when the published model name differs, including HybMNet for screening mammography (Chen et al., 28 Apr 2025), GLIMS for 3D brain tumor segmentation (Yazıcı et al., 2024), CAET-SWin for lung adenocarcinoma invasiveness prediction (Khademi et al., 2022), SDAH-UNet for medical image segmentation (Wang et al., 2023), and gSwin for general vision tasks (Go et al., 2022). The terminology is not uniform, however: the remote sensing model titled HSANet is explicitly a Hybrid Self-Cross Attention Network and not a Swin-based architecture (Han et al., 21 Apr 2025).

1. Terminology, scope, and naming ambiguity

The phrase “Hybrid Swin Attention Network” is best understood as an architectural category rather than a single canonical model. Across the cited papers, “hybrid” consistently denotes the combination of Swin-style windowed attention with a second mechanism that compensates for one of Swin’s practical limitations: weak fine-grained locality, inefficient dense upsampling, incomplete temporal modeling, or insufficient cross-scale aggregation (Chen et al., 28 Apr 2025, Yazıcı et al., 2024, Liu et al., 8 Sep 2025, Khademi et al., 2022, Wang et al., 2023, Go et al., 2022).

A recurrent source of confusion is that published names do not align perfectly with this conceptual category. HybMNet is not called HSANet in its paper, yet it is described as fitting “the concept of a Hybrid Swin Attention Network” because it combines a Swin-T backbone, saliency-driven ROI selection, a ResNet-18 local branch, and end-to-end fusion (Chen et al., 28 Apr 2025). GLIMS is likewise presented as an “attention-guided lightweight multi-scale hybrid network,” but its combination of 3D Swin blocks with convolutional multi-scale modules and attention-refined skips is explicitly mapped to HSANet in the supplied details (Yazıcı et al., 2024). By contrast, “HSANet: A Hybrid Self-Cross Attention Network for Remote Sensing Change Detection” does not use Swin Transformer, shifted windows, or a Swin-like hierarchy; the paper states that calling it a “Hybrid Swin Attention Network” is a misnomer (Han et al., 21 Apr 2025).

Model / term	Task	Hybridization pattern
HSANet	LDCT and PET denoising	Swin encoder–decoder + residual CNN + ESGA/EPGA + HIC
HybMNet	Mammography classification	Swin-T saliency branch + ResNet-18 ROI branch + fusion
GLIMS	3D brain tumor segmentation	DMSF/DMSU CNN modules + deep 3D Swin blocks + CSAB
CAET-SWin	LUAC prediction	Temporal CAE-Transformer path + spatial Swin path
SDAH-UNet	2D medical segmentation	SDMSA attention branch + parallel convolution branch
gSwin	Vision backbone	Swin hierarchy + Window-SGU gated MLP token mixing

This variation in naming suggests that HSANet is most usefully treated as a design pattern centered on Swin attention plus an explicitly non-Swin complementary pathway or module.

2. Core architectural pattern

The common substrate of HSANet-like systems is the Swin Transformer principle: hierarchical feature extraction with local self-attention inside non-overlapping windows, alternation between windowed MSA and shifted-window MSA, and progressive spatial downsampling through patch merging or an equivalent pyramid mechanism. In mammography, HybMNet uses a Swin-T backbone on $1024\times1024$ inputs with patch size $p = 4$, yielding $256\times256$ tokens, and applies attention within $16\times16$ token windows with shifted windows for cross-window interaction (Chen et al., 28 Apr 2025). In GLIMS, Swin blocks are inserted into the deeper encoder, bottleneck, and decoder of a full 3D U-Net-shaped network using $2\times2\times2$ patch tokenization and 3D shifted windows (Yazıcı et al., 2024).

Hybridization enters at the point where Swin alone is considered insufficient for the target task. HybMNet delegates high-saliency ROIs to a ResNet-18 branch to recover lesion-scale detail that may be diluted in global mammographic context (Chen et al., 28 Apr 2025). GLIMS uses DMSF and DMSU modules built from DACB blocks so that local multi-scale aggregation remains available at high resolutions, while CSAB-gated skip connections refine encoder features before decoder fusion (Yazıcı et al., 2024). The denoising HSANet inserts residual convolutional blocks in both encoder and decoder paths, replaces the Swin MLP with ESGA, refines skip connections through EPGA, and substitutes standard patch expansion with HIC to reduce overfitting to high-frequency noise (Liu et al., 8 Sep 2025).

Other instantiations extend the same hybrid logic in different directions. CAET-SWin splits the problem into temporal and spatial pathways: a CAE-Transformer path models inter-slice relations from per-slice latent codes, whereas a Swin-B path captures within-slice spatial structure from $224\times224$ nodule patches; both outputs are fused by fully connected layers (Khademi et al., 2022). SDAH-UNet places a Swin Deformable Multi-Head Self-Attention branch in parallel with a depthwise convolution branch inside each SDAPC block, thereby coupling deformable attention-based focalization with local texture recovery (Wang et al., 2023). gSwin preserves Swin’s four-stage hierarchical structure and shifted-window processing but replaces the MHSA token-mixing sublayer with Window-SGU, a gated MLP mechanism operating per local window (Go et al., 2022).

Across these systems, the unifying architectural claim is not merely “Transformer + CNN,” but a more specific strategy: Swin provides bounded-complexity contextualization, and the auxiliary branch or module restores information that a purely windowed Transformer may not represent optimally.

3. Attention mechanisms, fusion operators, and localization strategies

The mathematical backbone of most HSANet variants is standard Swin-style local attention with relative positional bias:

$$\mathrm{Attention}(Q, K, V)=\mathrm{softmax}\left(\frac{QK^T}{\sqrt{d}}+B_{\mathrm{rel}}\right)V.$$

This form appears directly in HybMNet and is structurally mirrored in GLIMS, the denoising HSANet, and CAET-SWin’s Swin branch (Chen et al., 28 Apr 2025, Yazıcı et al., 2024, Liu et al., 8 Sep 2025, Khademi et al., 2022).

What differentiates HSANet variants is how attention outputs are operationalized. In HybMNet, a $1\times1$ convolution plus sigmoid converts the Swin feature map into a saliency map $S$, and a greedy top-$k$ selection extracts $p = 4$0 ROI patches by maximizing the summed saliency within each candidate patch. The Swin branch yields a global prediction $p = 4$1 via GAP over the saliency map, the CNN branch produces $p = 4$2, and the fusion head combines global and local descriptors:

$p = 4$3

This makes the attention map not merely explanatory but causal in region proposal (Chen et al., 28 Apr 2025).

GLIMS uses attention differently. Its Swin blocks perform 3D windowed contextualization at reduced spatial scales, while CSAB applies channel attention followed by spatial attention to skip features before decoder fusion. The resulting encoder–decoder interaction is therefore attention-guided at two levels: Transformer attention for global context in coarse resolutions, and CBAM-like gating for skip refinement (Yazıcı et al., 2024).

The denoising HSANet introduces two lightweight attention modules. ESGA performs sequential channel and spatial attention and replaces the MLP inside each Swin block, while EPGA applies parallel channel and spatial attention to skip-connection features. The paper specifies ESGA’s channel reweighting as

$p = 4$4

and its spatial term uses $p = 4$5 convolutions with periodic shuffling and permutation operators to broaden effective receptive field with low parameter cost (Liu et al., 8 Sep 2025).

SDAH-UNet further modifies Swin attention through deformable sampling. For each window-head sub-feature, queries generate offsets through $p = 4$6, bilinear interpolation samples deformed keys and values, and the attention output becomes

$p = 4$7

This mechanism is explicitly tied to the paper’s notion of “model focalization,” in which deformation points cluster around the target anatomy or lesion (Wang et al., 2023).

A notable counterexample is the remote sensing HSANet, whose HSC-AFM uses self-attention and cross-attention over flattened multi-scale tokens from two dates. Because it lacks window partitioning and shifted windows, it should be classified as a hybrid attention network rather than a Hybrid Swin Attention Network in the strict sense (Han et al., 21 Apr 2025).

4. Training paradigms and supervision strategies

HSANet-like models do not share a standardized optimization regime; instead, supervision is tightly coupled to task structure. HybMNet is distinguished by a two-stage learning procedure. First, EsViT pretrains Swin-T on mammograms using a non-contrastive distillation framework with teacher–student Transformers, initialized from an SSL-pretrained ImageNet-1K Swin-T and adapted to larger windows by bicubic interpolation. Second, downstream end-to-end training optimizes the Swin branch, CNN branch, fusion head, and saliency sparsity jointly:

$p = 4$8

with $p = 4$9 in experiments (Chen et al., 28 Apr 2025).

GLIMS uses hierarchical deep supervision across four decoder heads. Each head employs a combined Dice and cross-entropy loss, and the total deep-supervision objective is

$256\times256$0

This supervision is designed to stabilize training and improve multi-resolution learning in a 3D setting (Yazıcı et al., 2024).

The denoising HSANet uses modality-specific losses. For LDCT, training combines MAE and Sobel loss:

$256\times256$1

whereas LDPET training uses MAE only because the Sobel term did not converge for PET (Liu et al., 8 Sep 2025).

CAET-SWin separates optimization across components. The CAE is pre-trained with MSE, fine-tuned partially, the temporal Transformer is trained with Adam, label smoothing $256\times256$2, and cross-entropy, the Swin-B path is fine-tuned with AdamW and cross-entropy, and the fusion head is trained separately with Adam and cross-entropy (Khademi et al., 2022). SDAH-UNet uses the sum of Dice loss and cross-entropy loss for segmentation (Wang et al., 2023). The remote sensing HSANet uses Dice loss only:

$256\times256$3

This diversity in training design indicates that HSANet is not defined by a single objective function but by a recurrent coupling of Swin attention with a task-specific complementary mechanism (Han et al., 21 Apr 2025).

5. Representative applications and reported performance

Reported results span classification, segmentation, denoising, and general-purpose vision. These metrics are task-specific and therefore not directly comparable across papers.

System	Application	Reported result
HybMNet_ssl	Screening mammography	AUC 0.864 on CMMD; 0.889 on INbreast
GLIMS	BraTS 2023 segmentation	Validation lesion-wise Dice 89.09 WT, 84.67 TC, 82.15 ET
HSANet	LDCT denoising	PSNR 32.21 ± 0.029; SSIM 0.90 ± 5.90e−4; RMSE 10.30 ± 0.040
HSANet	LDPET denoising	PSNR 34.034; SSIM 0.969; RMSE 5.928
CAET-SWin	LUAC invasiveness prediction	Accuracy 82.65%; sensitivity 83.66%; specificity 81.66%
SDAH-UNet	ACDC segmentation	AVG DSC 92.23; HD95 1.22
gSwin-T	ImageNet-1K	Top-1 81.71%

In screening mammography, HybMNet_ssl outperformed GMIC and GMIC_fc on both CMMD and INbreast. On CMMD test, HybMNet_ssl achieved AUC 0.864, Accuracy 0.797, and F1 0.790, compared with HybMNet at AUC 0.856, Accuracy 0.755, and F1 0.717; the paired t-test for AUC gave $256\times256$4. On INbreast test, HybMNet_ssl reached AUC 0.889, Accuracy 0.830, and F1 0.719, with the AUC comparison to HybMNet yielding $256\times256$5 (Chen et al., 28 Apr 2025).

In 3D brain tumor segmentation, GLIMS exceeded nnU-Net under matched training conditions in legacy Dice, averaging $256\times256$6 versus $256\times256$7. On the BraTS 2023 online validation set, it attained lesion-wise Dice scores of 89.09 for WT, 84.67 for TC, and 82.15 for ET, with average 85.30. On the blinded test set, it achieved 85.97 for WT, 82.90 for TC, and 83.67 for ET, with average 84.18 (Yazıcı et al., 2024).

In simultaneous low-dose denoising, the explicitly named HSANet achieved the best quantitative results among the listed baselines on both modalities. For LDCT, it reported PSNR $256\times256$8, SSIM $256\times256$9, RMSE $16\times16$0, GPU memory 5.4 GB, and 0.61 M parameters. For LDPET at DRF=100, it reported PSNR 34.034, SSIM 0.969, RMSE 5.928, GPU 5.4 GB, and 0.61 M parameters (Liu et al., 8 Sep 2025).

In lung cancer malignancy prediction, CAET-SWin reached accuracy $16\times16$1 with 95% CI $16\times16$2, sensitivity $16\times16$3, and specificity $16\times16$4 using 10-fold cross-validation on an in-house cohort of 114 pathologically proven sub-solid nodules. The hybrid exceeded both the Swin-only model, which achieved $16\times16$5 accuracy, and the best CAET-only variant, which achieved $16\times16$6 accuracy with GMP (Khademi et al., 2022).

In 2D medical segmentation, SDAH-UNet achieved AVG DSC 92.23 and HD95 1.22 on ACDC, and AVG DSC 86.90 and HD95 3.65 on BraTS2020, with reported statistical significance of $16\times16$7 versus the listed baselines (Wang et al., 2023). In general vision, gSwin-T reached $16\times16$8 top-1 on ImageNet-1K, box-AP 50.97 and mask-AP 44.16 on COCO, and mIoU 46.1 on ADE20K, while using fewer parameters and FLOPs than Swin-T (Go et al., 2022).

6. Limitations, reproducibility, and open questions

Several limitations recur across HSANet-like systems. HybMNet reports that transfer from CMMD-pretrained SSL to INbreast is limited by equipment and protocol differences, and the paper does not report runtime, memory footprint, or detailed per-stage Swin hyperparameters beyond window size and patch partitioning (Chen et al., 28 Apr 2025). GLIMS reports 47.16M parameters and 72.30G FLOPs, but also notes that 3D Swin blocks still incur nontrivial overhead and that post-processing materially improves lesion-wise metrics (Yazıcı et al., 2024). The denoising HSANet omits exact architectural hyperparameters such as window size, number of heads, embedding dimensions, and stage depths, and it does not report code availability, seeds, or data augmentation (Liu et al., 8 Sep 2025).

Reproducibility varies substantially. HybMNet code is “not publicly released” and is “available upon reasonable request to the corresponding author,” though PyTorch implementation details and hardware are specified (Chen et al., 28 Apr 2025). GLIMS provides code and specifies PyTorch 2.0.1 and MONAI 1.2.0 (Yazıcı et al., 2024). The remote sensing HSANet paper states that code will be open-sourced and provides a repository link, but because the architecture is not Swin-based, that implementation is relevant mainly as a boundary case in HSANet nomenclature (Han et al., 21 Apr 2025).

A broader conceptual limitation is terminological instability. Some papers are explicit HSANets, some are only retrospectively mapped to the concept, and some use “HSANet” for architectures that are not Swin-based. This suggests that the term is presently more useful as a descriptive umbrella for “Swin-centered hybrid attention architectures” than as a uniquely identifiable model family.

Future directions stated in the papers are correspondingly heterogeneous. Mammography work emphasizes larger and more diverse SSL pretraining corpora, stronger augmentation, and improved interpretability of saliency-linked radiological findings (Chen et al., 28 Apr 2025). GLIMS points to boundary-aware losses, uncertainty modeling, multi-domain adaptation, and lighter attention for 3D efficiency (Yazıcı et al., 2024). The denoising HSANet proposes cross-modality PET/CT fusion, adaptive window sizes, and kernelized or low-rank attention variants, while also identifying 3D deployment, inference speed, and uncertainty estimation as open clinical issues (Liu et al., 8 Sep 2025). Taken together, these directions indicate that HSANet research is converging less on a fixed blueprint than on a recurring engineering principle: Swin’s shifted-window hierarchy is retained for scalable contextual modeling, while hybrid modules are introduced wherever task-specific failure modes remain.