Laser & Optoelectronics Progress, Volume. 62, Issue 16, 1600002(2025)
Applications and Advancements of U-Net and Its Variants in Brain Tumor Image Segmentation
Figures & Tables(14)
Fig. 1. Comparison of tumor images with and without noise. (a) Noise-free image; (b) noisy image
Fig. 2. Brain tumor images with different MRI modals. (a) T1 weighted modal; (b) T2 weighted modal; (c) T1C modal; (d) FLAIR modal; (e) expert segmentation result
Fig. 3. Standard network architecture and network architecture with residual connections[27]. (a) Standard network architecture; (b) network architecture with residual connections
Fig. 8. Performances comparison of different algorithms on BraTS 2019 validation set
Table 1. Commonly used brain tumor image datasets
View tableTable 1. Commonly used brain tumor image datasets
Dataset Main application Applicable scenario Modal Image size Source BraTS Brain tumor segmentation， multimodal MRI data Tumor detection， segmentation model benchmarking T1、T1-Gd、T2、FLAIR 240 pixel ×240 pixel ×155 pixel （typical） https：//www.med.upenn.edu/sbia/brats2020/ TCIA General medical imaging， cancer research Cancer diagnosis， image segmentation， model development MRI 256 pixel ×256 pixel （variational） https：//www.cancerimagingarchive.net/ ISLES Stroke lesion detection， MRI segmentation Stroke detection， lesion localization， clinical study MRI 256 pixel ×256 pixel or 128 pixel ×128 pixel http：//www.isles-challenge.org/ RSNA Clinical brain MRI data， disease diagnosis Disease diagnosis， brain structure analysis MRI 512 pixel ×512 pixel （variational） https：//www.rsna.org/en Lower-Grade Glioma Lower-Grade Glioma segmentation， MRI data Lower-Grade Glioma research， tumor progression study MRI 240 pixel ×240 pixel ×155 pixel （typical） https：//www.med.upenn.edu/sbia/lgg-challenge/ CMB Dataset Microbleed detection， MRI data Microbleed detection， small lesion analysis MRI 256 pixel ×256 pixel （variational） https：//pubmed.ncbi.nlm.nih.gov/
Table 2. Commonly used brain tumor image segmentation indices and their expressions
View tableTable 2. Commonly used brain tumor image segmentation indices and their expressions
Index Formula Dice coefficient Sensitivity Specificity Precision
Table 3. Comparison of the improved U-Net mechanisms
View tableTable 3. Comparison of the improved U-Net mechanisms
Method Mechanism Advantage Limitation Algorithm Feature Network structure and feature optimization Multi-scale feature extraction Effectively capture tumor features at different scales， improve segmentation accuracy for small tumors and boundary regions High computational cost， complex model， which can burden training and inference processes RAPNet［ 22 ］Multi-scale， dilated convolutions， attention mechanism HA-RUnet［ 23 ］Attention mechanism， SE modules Ga-U-Net［ 24 ］Gabor convolutions， attention mechanisms U-shaped encoder-decoder model［ 25 ］Compact split attention， enhanced feature extraction Residual and skip connection Increase network depth and width while avoiding gradient vanishing problems， improve feature extraction and classification accuracy May not fully solve gradient vanishing issues， limited ability to capture small tumor boundaries IRDNU-Net［ 27 ］Residual-Inception modules Residual learning U-Net［ 29 ］Residual learning， feature extraction dResU-Net［ 30 ］The training process is optimized and feature extraction is enhanced by jump connections between residual and convolutional blocks MMGAN［ 31 ］Residual learning， reduced parameters Lightweight design Reduce model parameters， lower computational cost， and improve robustness and efficiency May lose detail in complex tumor areas， perform poorly on highly complex data SEDNet［ 36 ］Hierarchical convolution， feature learning， robustness， efficiency， optimized architecture， segmentation， fewer parameters GA-UNet［ 38 ］Lightweight design， GhostV2 bottleneck， attention module Contextual information and attention mechanism Context information and attention mechanism Improve tumor localization and boundary recognition， reduce background interference Limited ability to capture complex boundaries and fine details， performance may be affected by background complexity MMS-Net［ 39 ］Triple attention modules， multi-modal MRI segmentation TDPC-Net［ 40 ］3D attention， decoupled convolution units Dual attention U-Net［ 43 ］Dual attention mechanism， iterative feature aggregation 3D U-Net with attention［ 44 ］Residual network， attention mechanisms， adaptive learning Transformer fusion Strong global context capture ability， improved segmentation accuracy， especially in multi-modal data High computational and memory overhead， long training time， leading to lower processing efficiency UNETR［ 53 ］Transformer encoder， global context modeling TransMVU［ 55 ］Transformer + U-Net， multi-view performance Swin-UNet［ 58 ］Swin-Transformer， global context modeling Spatial pyramid pooling Expand the receptive field and allow for multi-scale feature extraction， preserve fine details Performance may degrade with very high-resolution images， less effective in very fine details Attention-UNet with ASPP［ 61 ］Attention mechanism， ASPP， multi-scale feature extraction， expanded receptive fields， preserved fine details， and improved segmentation accuracy Training strategy and performance improvement Training strategy and performance improvement Significantly enhance segmentation of small tumors and imbalanced data， improve boundary and overlap accuracy Less effective for large and complex tumors， may not handle large tumor structures well Weighted loss + Dice loss［ 62 ］Generalized Dice loss， attention mechanism MUNet［ 63 ］mIoU loss， Dice loss， boundary loss， small tumor regions， overlap， similarity SBTC-Net［ 64 ］Transfer learning， segmentation and classification
Table 4. Comparison of pre-training methods for U-Net
View tableTable 4. Comparison of pre-training methods for U-Net
Method Mechanism Advantage Limitation Algorithm Feature Multi-modal fusion Utilize complementary information from different modals （MRI， CT） for enhanced feature extraction Improve segmentation accuracy by integrating information from different modals Potential risk of insufficient feature fusion ACMINet［ 68 ］Cross-modal feature alignment， accurate tumor segmentation， volume brain tumor focus F2 Net［ 69 ］Effective feature fusion strategies that allow information from different modals to complement each other CKD-TransBTS［ 70 ］By integrating the radiologist's clinical experience with the model's structure and combining a two-branch hybrid encoder with the cross-attention mechanism SDPN［ 71 ］Lightweight dual-path network， multi-spectral attention， and tensor ring decomposition GAN and Image Augmentation Combine U-Net with GANs （cGAN） to generate realistic tumor images and improve segmentation Enhance segmentation precision and boundary details， improve model robustness to different tumor shapes Stability issues with GAN parameters， high complexity Adversarial U-Net［ 73 ］Generative adversarial network integration， U-Net as generator， enhanced segmentation image quality Pix2Pix［ 74 ］Conditional GAN model， image-to-image translation， high-quality image generation， effective for paired datasets CycleGAN［ 75 ］Unpaired image-to-image translation， cycle-consistency loss， no need for paired datasets Self-supervised and semi-supervised learning Use self-supervised learning to pre-train models and semi-supervised learning with few labeled data Enhance model performance with fewer labeled images， improve robustness and adaptability Self-supervised learning still has limitations in complex segmentation tasks SSW-AN［ 80 ］Self-supervised learning， adjustable activation function， dynamic weight attention module， low/high-frequency channel separation Similarity-based algorithm［ 81 ］Semi-supervised learning， similarity constraints， effective with limited labeled data， improved generalization Real-time segmentation and clinical application Combine GAN， self-supervised learning， and efficient architectures to support fast inference and clinical applications Increase segmentation speed while maintaining accuracy， suitable for real-time clinical use Dependent on high-quality input data， limited application in low-quality data scenarios Cascaded method［ 82 ］Two-stage approach， WT region detection， fine-grained sub-region segmentation TransU2-Net［ 84 ］Transformer module integration， lightweight design， efficient feature extraction EfficientNet-U-Net［ 85 ］EfficientNet integration， optimized computational efficiency， high segmentation accuracy
Table 5. Performance comparison of different types of U-Net in brain tumor image segmentation tasks
View tableTable 5. Performance comparison of different types of U-Net in brain tumor image segmentation tasks
Category Advantage Limitation Methodology Dataset Precision /% Others TC ET WT U-Net+
attention mechanism
Dynamically highlight key features， improve local segmentation precision， excel in enhancing tumor regions Sensitive to parameter initialization， unstable training， increase model complexity and inference time MMS-Net BraTS 2019 84.5 83.0 90.6 — TDPC-Net BraTS 2019 84.1 79.1 90.2 — DAMIA-U-Net BraTS 2019 71.5 71.3 80.5 — RF-ATT-U-Net BraTS 2018 86.0 79.0 86.0 — GCA-U-Net BraTS 2021 89.7 85.6 91.0 Average Hausdorff distance： 9.12 mm MUNet BraTS 2018 81.5 81.5 90.1 Hausdorff95： 6.152
4.389 6.243
U-Net+
skip connection mechanism
Preserve spatial information and improve gradient flow， enhance segmentation precision by combining low-level and high-level features Increase computational complexity and may struggle with processing complex structures mResU-Net BraTS 2021 92.9 89.7 92.8 — DSDU-Net — — — — Preserve low-level features， multi-scale patch extraction for tumor region separation ResNet with Attention Gates — — — — Combine ResNet with U-Net， attention gates improve
tumor localization precision
Maximized U-Net — — — — Enhance sensitivity to multi-scale information， improve feature expression ability U-Net+
residual linkage mechanism
Mitigate gradient vanishing， enhance deep feature representation， achieve more accurate segmentation Highly dependent on data generalization， increased model complexity and parameters dResU-Net — — — — Residual network as encoder to mitigate gradient vanishing；
U-Net decoder improves segmentation performance
RAPNet BraTS 2019 85.2 85.2 85.2 — U-Net+
dense connection mechanism
Enhance feature reuse and gradient flow， improve segmentation performance by allowing the network to efficiently learn from earlier layers Increase model complexity and computational cost Improved U-Net — — — — Enhance segmentation precision and robustness by adjusting network depth and channel count Data Augmentation Strategy with U-Net — — — — Combine effective data augmentation to handle data diversity in brain tumor segmentation U-Net+
Transformer
Capture long-range dependencies and global features， excel in complex tasks with superior performance High computational complexity， require strong hardware support， unsuitable for real-time applications DiffSwinTr — 79.9 83.07 85.38 Average Hausdorff distance： 3.361 mm； 3.334 mm； 2.975 mm（95%） U-Net+
GAN
Generate high-quality synthetic data， alleviate data scarcity， excel in small-sample segmentation tasks Unstable training， heavily rely on parameter tuning， segmentation performance depend on generated data quality Swin-UNet — 78.0 78.0 89.0 — Inception-UDet — 81.0 77.0 89.0 — ESGan BraTS 2013 82.0 76.0 89.0 — MMGan BraTS 2018 86.0 78.0 87.0 — U-Net+
multimodal
Combine multimodal information to enhance segmentation precision Rely on high-quality multimodal data， which can be difficult to acquire and align F2 Net — — — — Significant improvement in technology CKD-TransBTS — 90.2 88.5 93.3 Sensitivity： 90.6%（TC） 90.0%（ET） 90.3%（WT） TransDoubleU-Net BraTS 2019/2020 — — — Promising results in
MRI
modal grading
U-Net++
EfficientNet
LGG — — — Average Dice coefficient： 91.8% GA-UNet ISIC-2018 — — — F1-score： 89.6% LGG — — — F1-score： 89.6% U-Net+DSM BraTS 2019 — — — Sensitivity： 98.59%
Specificity： 98.64%
Precision： 98.64%
Average Dice coefficient： 98.02%
SAU-Net BraTS 2017 — — — Dice loss： 54.8%（TC）
56.3%（ET）
68.6%（WT）
TransU2-Net BraTS 2021 — — — Average Dice coefficient： 88.17% HF-U-Net BraTS 2018 — — — Average Dice coefficient： 91.34%；
Hausdorff distance： 3.74 mm
U-Net-EfficientNet BraTS 2019 81.5 82.0 89.0 — ETumorNet-YOLOv8-U-Net TCIA — — — Precision： 98.6%；
Recall： 95.2%；
F1-score： 96.3%；
Specificity： 89.1%
Table 6. Performance comparison of U-Net with different training strategies on BraTS 2020
View tableTable 6. Performance comparison of U-Net with different training strategies on BraTS 2020
Category Advantage Limitation Methodology Precision /% Others TC ET WT U-Net+
attention mechanism
Multi-scale feature fusion， enhanced boundary preservation High computational complexity HA-RUnet 81.30 78.70 86.70 Sensitivity： 83.0%（TC）；
88.0%（ET）；
93.0%（WT）
Hierarchical feature extraction enhances tumor localization Require extensive hyperparameter tuning MUNet 82.30 83.50 91.50 Hausdorff95： 6.437
2.421
3.755
U-Net+
residual linkage mechanism
Cross-resolution feature fusion enhances robustness Challenging deployment on resource-limited systems MCG-CR3D-U-Net 89.90 84.90 91.70 Hausdorff distance： 9.203 mm（TC）
12.171 mm（ET）
6.021 mm（WT）
U-Net+
GAN
Generative augmentation mitigates data scarcity Training instability due to GAN optimization challenges Vit+GAN — — — Precision： 97.65%；
Sensitivity：97.70%
Lightweight improvement Depthwise separable convolutions minimize parameters Lower precision compared to heavy-weight models SEDNet 90.26 94.51 93.08 Hausdorff distance： NTC （Nearest Point Correspondence） is 0.704 mm， ED （Euclidean Distance） is 1.286 mm， ET（Error Threshold） is 0.776 mm U-Net+
transfer learning
Incorporate spatial and boundary attention for precise segmentation Potential overfitting on small datasets SBTC-Net — — — Precision： 99.97%
Tools
Get Citation
Copy Citation Text
Nan Wang, Hua Wang, Dejian Wei, Liang Jiang, Peihong Han, Hui Cao. Applications and Advancements of U-Net and Its Variants in Brain Tumor Image Segmentation[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1600002
Paper Information
Category: Reviews
Received: Dec. 6, 2024
Accepted: Mar. 12, 2025
Published Online: Aug. 8, 2025
The Author Email: Hui Cao (caohui63@163.com)
CSTR:32186.14.LOP242385
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20