Skip to main content
Have a personal or library account? Click to login
Swinnetplus: A Novel Deep Learning Approach Using Swin Transformer For Enhancing Brain Tumor Segmentation Cover

Swinnetplus: A Novel Deep Learning Approach Using Swin Transformer For Enhancing Brain Tumor Segmentation

By:  and    
Open Access
|Sep 2026
Read online

Figures & tables

Figure 1.

Illustration depicting various sub-regions for brain tumor segmentation [1]

Figure 2.

SwinNetPlus Architecture

Figure 3.

The Swin Transformer architecture comprises distinct components. (a) Self-attention computations within adjacent layers l and l +1 utilize both regular and shifted windows. (b) Further insights into the intricacies of the Swin Transformer are provided

Figure 4.

Swin Transformer with ELSA Block [20]

Figure 5.

DCML Module

Figure 6.

The architecture of the SLAM module

Figure 7.

Channel Squeeze and Spatial Excitation (sSE[21])

Figure 8.

Illustration depicting various modalities of BraTS 2021[1]

Figure 9.

Segmentation results of the proposed method: a. FLAIR, b. T1ce, c. proposed method, and d. ground truth segmentation

Table 1.

Summary of Key Related Works and Limitations

Author(s)MethodKey ContributioneLimitation
Long et al. [2]FCNEnd-to-end semantic segmentationPoor boundary localization
Ronneberger et al. [3]U-NetEncoder–decoder with skip connectionsLimited global context
Zhou et al. [4]UNet++Nested skip connectionsIncreased complexity
Çiçek et al. [5]3D U-NetVolumetric segmentationHigh memory usage
Dosovitskiy et al. [8]ViTGlobal self-attentionComputationally expensive
Liu et al. [9]Swin TransformerEfficient hierarchical attentionWeak local boundary sensitivity
Chen et al. [13]TransUNetHybrid CNN–TransformerFeature inconsistency
Hatamizadeh et al. [14]UNETRTransformer encoder for 3D dataRequires large datasets
Jiang et al. [16]SwinBTSSwin-based brain tumor segmentationLimited local attention
Ghazouani et al. [21]Swin + ELSAEnhanced local attentionInadequate multiscale fusion
Xing et al. [17]NestedFormerModality-aware transformerHigh computational cost
Table 2.

Segmentation outcomes observed on the BraTS 2021 validation dataset

MethodDice Score (%)95% Hausdorff Dist. (mm)
ETTCWTAVG.ETTCWTAVG.
UNETR[14]58.576.178.971.169.358.848.268.81
3DUNet[5]83.9385.1189.1585.974.7311.9212.2610.26
NestedFormer[17]8086.49286.135.265.314.565.04
SwinBTS[16]83.2184.7591.8386.5916.0315.513.658.54
SegTransVAE[22]85.4890.5292.689.532.893.575.844.1
Swin UNETR[18]85.888.592.688.966.013.775.835.2
Our Method90.1294.1993.7792.697.324.44.315.34
Table 3.

Ablation study of SwinNetPlus on the BraTS 2021 validation dataset

Model VariantDCMLSLAMMult. FusionDice Score (AVG) ↑95% HD (AVG) ↓
Baseline ( without DCML/SLAM)✗✗✗87.156.93
Baseline + DCML✓✗✓88.766.15
Baseline + SLAM✗✓✗89.215.94
Baseline + DCML + SLAM (Additive Fusion)✓✓✗90.035.59
Full Model (SwinNetPlus)✓✓✓92.695.34
Figure 10.

Ablation Study Results of SwinNetPlus on the BraTS 2021 Validation Dataset

DOI: https://doi.org/10.14313/jamris-2026-049 | Journal eISSN: 2080-2145 | Journal ISSN: 1897-8649
Language: English
Page range: 183 - 194
Submitted on: Jun 6, 2025
Accepted on: Feb 2, 2026
Published on: Sep 30, 2026
Published by: Łukasiewicz Research Network – Industrial Research Institute for Automation and Measurements PIAP
In partnership with: Paradigm Publishing Services
Publication frequency: 4 issues per year

© 2026 Vikash Verma, Pritaj Yadav, published by Łukasiewicz Research Network – Industrial Research Institute for Automation and Measurements PIAP
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.