SCANet: Versatile Attention-Based Architectures

• SCANet is a collection of neural network architectures characterized by attention mechanisms, contextual aggregation, and task-specific adaptations.
• It encompasses models applied in medical imaging, image restoration, remote sensing, video retrieval, and tracking, each optimized for its unique data inputs.
• SCANet variants demonstrate enhanced performance and efficiency compared to traditional models, though they differ in design, fusion strategies, and computational demands.

SCANet is a reused neural-network designation rather than a single canonical architecture. In the arXiv literature, it denotes multiple task-specific models spanning acute ischemic stroke prognosis, image dehazing, real-image denoising, cloud segmentation, building footprint extraction, weakly supervised video moment retrieval, LiDAR panoptic segmentation, LEGO assembly correction, and multimodal underwater tracking. Across these works, the shared lexical motif is usually attention, cross-modal interaction, or context aggregation, but the architectural semantics of the acronym vary substantially by domain and should not be conflated (Zhang et al., 2023, Guo et al., 2023, Yoon et al., 2023, Wang et al., 28 Jul 2025, Wan et al., 2024).

1. Nomenclature and scope

The acronym has been expanded in multiple incompatible ways. In addition, several closely related names—such as SaNet, SCanNet, and MS-SCANet—are adjacent rather than identical usages, and in some cases are explicitly described as being loosely referred to as “SCANet” in other contexts (Wang et al., 2021, Ding et al., 2022, Mithila et al., 3 Feb 2026).

Expansion or form	Task domain	arXiv id
Spatial Cross Attention Network	Thrombectomy recanalization prediction from CT/CTA	(Zhang et al., 2023)
Self-Paced Semi-Curricular Attention Network	Non-homogeneous image dehazing	(Guo et al., 2023)
Structure-preserving Complementarity Attention Network	Real image denoising/restoration	(Zhang et al., 2022)
Scene Complexity Aware Network	Weakly-supervised video moment retrieval	(Yoon et al., 2023)
Segregation and Context Aggregation Network	Real-time cloud segmentation	(Li et al., 19 Apr 2025)
Split Coordinate Attention Network	Building footprint extraction	(Wang et al., 28 Jul 2025)
Sparse Cross-scale Attention Network	LiDAR panoptic segmentation	(Xu et al., 2022)
Self-Correct Assembly Network	LEGO assembly error correction	(Wan et al., 2024)

A recurrent misconception is that SCANet refers to a single lineage comparable to ResNet or U-Net. The literature instead shows a naming pattern in which separate research groups reuse the acronym for architectures optimized around different inductive biases, losses, and output spaces. Closely related but distinct names include SCanNet, the Semantic Change Network for remote-sensing semantic change detection, and MS-SCANet, a Multi-Scale Spatial Channel Attention Network for no-reference image quality assessment (Ding et al., 2022, Mithila et al., 3 Feb 2026).

2. Spatial Cross Attention Network in stroke imaging

In medical imaging, SCANet denotes a Spatial Cross Attention Network for predicting mechanical thrombectomy recanalization success directly from pre-treatment CT and CTA in acute ischemic stroke due to large-vessel occlusion (Zhang et al., 2023). The task is framed as binary classification of favorable versus unfavorable post-procedure reperfusion, using the modified Treatment in Cerebral Ischemia criterion, with favorable outcome defined as mTICI 2c or greater and unfavorable outcome as mTICI less than 2c. The model is fully automatic and does not require manual clot segmentation.

The study retrospectively collected single-center data from 2012–2019. From 254 eligible patients, 177 remained for modeling after excluding 69 patients with missing CT or CTA and 8 patients with unclear stroke location; the target was approximately balanced. Preprocessing used brain extraction and affine/nonlinear registration to a CT template in MNI space, enabling slice-wise modeling. The network operated in a 2D slice-wise fashion with slices resized to $224 \times 224$, and the first convolutional block received an input described as $26 \times 224 \times 224$ (Zhang et al., 2023).

Architecturally, the model combines a ResNet34-based CNN with two transformer modules: a Spatial Attention Transformer (SAT) for within-slice localization and a Cross Attention Transformer (CAT) for weighting slices within a neighborhood. SAT follows the standard attention form

$$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$

where the tokens are spatial patches within a slice. CAT applies the same formulation across slice tokens, thereby localizing along the cranio-caudal axis. Final branch outputs are aggregated and passed through a weighted softmax layer, and the objective is binary cross-entropy loss (Zhang et al., 2023). The most plausible interpretation presented for CT and CTA fusion is early fusion at input level, likely as separate channels, though that fusion detail is not given as an explicit architectural statement.

Performance was reported by cross-validation. SCANet achieved ROC-AUC $0.7732 \pm 0.039$, accuracy $0.7523 \pm 0.042$, precision $0.7424 \pm 0.145$, sensitivity $0.8250 \pm 0.174$, and specificity $0.6905 \pm 0.215$. This exceeded a radiomics model with ROC-AUC $0.6859 \pm 0.043$ and a standard ResNet34 baseline with ROC-AUC $0.5840 \pm 0.036$, and also improved on previously published fully automatic and semi-automated approaches reporting 0.717 and $26 \times 224 \times 224$0 ROC-AUC, respectively (Zhang et al., 2023). The principal limitations were the single-center cohort, the modest sample size of 177 patients, and the authors’ concern that preprocessing adapted for CT might not optimally preserve high-resolution CTA detail.

3. Restoration-oriented SCANets

In non-homogeneous dehazing, SCANet denotes a Self-Paced Semi-Curricular Attention Network with two sub-networks: an Attention Generation Network (AGN) and a Scene Reconstruction Network (SRN) (Guo et al., 2023). The AGN is composed of stacked Dual Attention Units (DAUs), each combining channel attention with Multi-Scale Pixel Attention using dilated convolutions at rates $26 \times 224 \times 224$1. The SRN is an encoder–decoder with residual blocks and two deformable convolution layers. A luminance-derived attention ground truth $26 \times 224 \times 224$2, computed from Y-channel differences in YCbCr space, supervises the learned attention map $26 \times 224 \times 224$3, and during the first 25% of training epochs the final attention is blended by

$26 \times 224 \times 224$4

Feature modulation in SRN uses

$26 \times 224 \times 224$5

On NTIRE2020, NTIRE2021, and NTIRE2023, the full model reported average PSNR 20.37 and average SSIM 0.6933, with individual results of 19.52 dB / 0.6488 on NTIRE2020, 21.14 dB / 0.7694 on NTIRE2021, and 20.44 dB / 0.6616 on NTIRE2023 (Guo et al., 2023).

For real image denoising, SCANet denotes a Structure-preserving Complementarity Attention Network built as a dual-branch U-Net-style CNN (Zhang et al., 2022). Its core Complementary Attention Module (CAM) combines a Dense Module with spatial and channel attention and a Sparse Module that uses cheap linear operations to generate sparse feature maps from a smaller base set. A second, gradient-based branch predicts clean gradient maps and provides structural priors back to the pixel branch. The total training objective combines an improved Charbonnier pixel loss with pixel-based and gradient-branch gradient losses. On SIDD, SCANet reported 39.60 dB PSNR and 0.969 SSIM; on DND, it reported 39.65 dB PSNR and 0.953 SSIM. The model required 66.58 GFLOPs on $26 \times 224 \times 224$6 inputs and had an inference runtime of 54.7 ms, compared with 335.01 GFLOPs and 121.3 ms for CycleISP (Zhang et al., 2022).

For ground-based atmospheric segmentation, SCANet denotes a Segregation and Context Aggregation Network for real-time binary sky/cloud segmentation (Li et al., 19 Apr 2025). Its decoder module, SCAM, uses the previous stage’s rough segmentation map to separate features into weighted sky and cloud branches, processes both with inverted residual blocks, and then aggregates them. The family includes SCANet-large, SCANet, and SCANet-lite. On SWINySEG, SCANet-large (4.29M) achieved accuracy 0.970 and MIoU 0.923, while SCANet-lite (90K) achieved 1390 fps in FP16 on a V100 GPU and accuracy 0.944 with MIoU 0.865. The paper also introduced a task-specific pre-training strategy that operates without ImageNet pre-training (Li et al., 19 Apr 2025).

4. Remote sensing and geospatial analysis

In building footprint extraction, SCANet denotes a Split Coordinate Attention Network formed by inserting the Split Coordinate Attention (SCA) module into a 2D CNN backbone and pairing it with a UNet++ decoder (Wang et al., 28 Jul 2025). SCA combines split-attention grouping with coordinate-aware pooling along height and width, using two spatial ranges of pooling kernels $26 \times 224 \times 224$7 and $26 \times 224 \times 224$8. It is designed to capture remote spatial interactions while preserving positional information that would be lost by global average pooling. With a ResNeSt-style encoder and UNet++, the model achieved IoU 91.61% on the WHU Building Dataset and 75.49% on the Massachusetts Building Dataset, outperforming SA- and CA-based variants while using 73.2M parameters in the UNet++ setting, equal to SA and lower than CA (Wang et al., 28 Jul 2025).

A related but distinct naming variant is SCanNet, the Semantic Change Network for semantic change detection in remote sensing (Ding et al., 2022). It combines a triple encoder–decoder CNN backbone with SCanFormer, a Semantic Change Transformer based on CSWin self-attention, and a semantic learning scheme that enforces spatio-temporal consistency through

$26 \times 224 \times 224$9

The method jointly reasons over semantic maps at $$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$0 and $$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$1 and a change branch, rather than treating semantic segmentation and binary change detection as loosely coupled outputs. On SECOND, SCanNet reported OA 87.86, mIoU 73.42, SeK 23.94, and $$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$2 63.66; on Landsat-SCD, it reported OA 96.26, mIoU 88.96, SeK 60.53, and $$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$3 89.27 (Ding et al., 2022).

The name SaNet also sits near this terminological cluster. In the remote-sensing segmentation literature it is explicitly described as being “sometimes loosely referred to as a ‘scale-aware CNN’/SCANet in other contexts” (Wang et al., 2021). SaNet is a fully convolutional semantic segmentation model for multi-resolution remote sensing imagery, built around a Spatial Feature Recalibration Module (SFRM) and a Densely Connected Feature Fusion Module (DCFFM). It reported 81.2 test mIoU on the cross-resolution LandCover.ai setup and mean OA 83.4 on both MSR Vaihingen and MSR Potsdam, with performance degradation under resolution reduction described as the slowest among compared models (Wang et al., 2021). This usage underscores that “SCANet” in remote sensing may denote either exact acronym matches or proximate naming variants with related attention semantics.

5. Retrieval and tracking architectures

In weakly supervised video moment retrieval, SCANet denotes a Scene Complexity Aware Network that replaces fixed proposal heuristics with complexity-adaptive proposal generation (Yoon et al., 2023). Scene complexity is defined as

$$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$4

where $$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$5 is the number of distinct scenes in a video inferred from associated queries in the dataset. That scalar conditions a learnable codebook vector $$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$6, the selected number of proposals $$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$7, proposal centers and widths, and a Flatten Gaussian mask. The training objective combines masked query reconstruction, masked video reconstruction, video-level contrastive loss, corpus-level contrastive loss, and a complexity-aware calibration factor. On Charades-STA, SCANet reported R@1, IoU=0.5 of 50.85 and R@1, IoU=0.7 of 24.07; on ActivityNet Captions, R@1, IoU=0.5 of 31.52; and on TVR, R@1, IoU=0.5 of 4.24. The method’s main conceptual claim is that proposal budgets should respond to scene complexity rather than video length (Yoon et al., 2023).

In underwater tracking, SCANet is an RGB–sonar tracker introduced together with the RGBS50 benchmark, which contains 50 sequences and 87,404 annotated bounding boxes (Li et al., 2024). The architecture is based on OSTrack but adds a Spatial Cross-Attention Module (SCAM) at layers 4, 7, and 10 to address the fact that RGB and sonar are time-aligned but spatially misaligned. For RGB and sonar token sequences $$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$8 and $$\text{Attention}(Q, K, V)=\text{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$9, cross-attention is computed with cross-modal queries and keys, and the paper replaces softmax by ReLU: $0.7732 \pm 0.039$0 Independent Global Integration Modules then refine each branch. Because no real RGB–sonar training corpus existed, the paper also proposed SRST, which synthesizes pseudo RGB–sonar pairs by converting RGB images into sonar-like saliency images and augmenting with SARDet-100K. SCANet reported RGB SR/PR/NPR = 71.2/75.6/78.0 and sonar SR/PR/NPR = 50.3/66.2/62.1, improving substantially over OSTrack256, particularly on sonar where the baseline achieved 40.9/54.6/51.6 (Li et al., 2024).

These two SCANets share a structural concern with latent correspondence under incomplete supervision, but the mechanism differs sharply. The retrieval model adapts temporal proposals to inferred scene multiplicity, whereas the tracker models cross-modal correspondences without assuming any pixel-wise or box-wise alignment.

6. 3D perception, robotic correction, and derivative variants

In LiDAR panoptic segmentation, SCAN denotes a Sparse Cross-scale Attention Network that operates on sparse voxels rather than dense BEV tensors (Xu et al., 2022). Its cross-scale module aligns multi-scale sparse voxel coordinates, applies global attention with 3D positional encoding, and then predicts a sparse, class-agnostic BEV centroid heatmap. The total loss is

$0.7732 \pm 0.039$1

combining centroid heatmap supervision, point-wise offset loss, semantic segmentation loss, and multi-scale sparse semantic supervision. On SemanticKITTI, SCAN reported PQ 61.5%, $0.7732 \pm 0.039$2, mIoU 67.7%, and 12.8 FPS; on nuScenes, it reported PQ 65.1% and mIoU 77.4% (Xu et al., 2022). Its design goal is to improve long-range geometric reasoning for surface-aggregated point clouds while avoiding time-consuming clustering.

In robotic assembly, SCANet denotes a Self-Correct Assembly Network introduced with the LEGO Error Correction Assembly Dataset (LEGO-ECA) (Wan et al., 2024). The paper defines the Single-Step Assembly Error Correction Task, in which an upstream assembler such as MEPNet has already produced a step result, and SCANet must classify each current-step component as correct, position error, rotation error, or both, then regress a corrected pose. LEGO-ECA was created by perturbing manual images with Gaussian noise before running MEPNet, yielding roughly 120,000 incorrectly assembled examples. SCANet uses a pair of manual-shape and assemble-shape branches, an assembly-difference extractor, a component pose encoder that fuses 3D voxel shape, normalized 6D pose, and per-component 2D appearance, a DETR-style transformer, and three heads for status, position, and rotation. When used to correct MEPNet on LEGO-ECA, it improved component pose accuracy from 38.58% to 49.31%, step pose accuracy from 28.66% to 36.96%, and reduced Chamfer distance from 28.98 to 12.61, with Correction Rate 19.31% and Misplacement Rate 6.35% (Wan et al., 2024).

A later derivative usage is MS-SCANet, a Multi-Scale Spatial Channel Attention Network for no-reference image quality assessment (Mithila et al., 3 Feb 2026). This 2026 model explicitly frames itself as extending a SCANet-like spatial–channel attention idea into a dual-branch transformer operating on 16×16 and 32×32 patches, with cross-branch attention and two consistency losses: $0.7732 \pm 0.039$3 Its reported main results were PLCC/SROCC 0.903/0.909 on KonIQ-10k, 0.968/0.964 on LIVE, 0.903/0.895 on LIVE-C, and 0.945/0.925 on CSIQ; with the full L1 + CB + AP configuration, the reported scores rose to 0.921/0.923 on KonIQ-10k and 0.972/0.973 on LIVE (Mithila et al., 3 Feb 2026). This usage is notable because it treats SCANet less as a fixed architecture than as a design family centered on joint spatial and channel attention.

Taken together, these models show that “SCANet” functions in contemporary deep-learning literature as a task-local label for architectures that privilege structured attention, cross-branch interaction, or explicit correction. The commonality is methodological rather than genealogical: each SCANet is tightly coupled to its own data representation, supervision regime, and error model, and direct comparisons across papers are meaningful only within the originating application domain.