Source-linked AI summary
SAR-U-Net: squeeze-and-excitation block and atrous spatial pyramid pooling based residual U-Net for automatic liver segmentation in Computed Tomography
Jinke Wang, Peiqing Lv, Haiying Wang, Changfa Shi
TL;DR
Manual liver segmentation from CT is labor-intensive, motivating accurate and robust automatic methods that remain challenging on difficult liver regions and boundaries. The paper proposes SAR-U-Net, combining SE attention, ASPP, residual learning, and weighted cross-entropy, and reports superior performance on LiTS17 and SLiver07 relative to related classic methods. The model is robust on small, discontinuous, and blurred-boundary liver cases, but its 2D design can miss z-axis context and may err near lesions or tumors.
Problem
Manual CT liver labeling is labor-intensive, while existing automatic methods remain insufficiently accurate and robust for clinical data and challenging liver cases.
Method
SAR-U-Net combines SE attention, ASPP-based multi-scale extraction, residual learning, and weighted cross-entropy for automatic liver CT segmentation.
Results
SAR-U-Net achieved superior quantitative performance on LiTS17 and SLiver07 relative to related classic methods, with robust handling of difficult segmentation cases.
Takeaways & Limitations
The method improves reported accuracy and robustness for small liver regions, discontinuous regions, and blurred liver boundaries.
Takeaways & Limitations
Because SAR-U-Net is 2D while CT images are 3D, it can miss z-axis context and produce large boundary errors around lesions or tumors.
Abstract
from arXiv · showhide
Background and objective: In this paper, a modified U-Net based framework is presented, which leverages techniques from Squeeze-and-Excitation (SE) block, Atrous Spatial Pyramid Pooling (ASPP) and residual learning for accurate and robust liver CT segmentation, and the effectiveness of the proposed method was tested on two public datasets LiTS17 and SLiver07. Methods: A new network architecture called SAR-U-Net was designed. Firstly, the SE block is introduced to adaptively extract image features after each convolution in the U-Net encoder, while suppressing irrelevant regions, and highlighting features of specific segmentation task; Secondly, ASPP was employed to replace the transition layer and the output layer, and acquire multi-scale image information via different receptive fields. Thirdly, to alleviate the degradation problem, the traditional convolution block was replaced with the residual block and thus prompt the network to gain accuracy from considerably increased depth. Results: In the LiTS17 experiment, the mean values of Dice, VOE, RVD, ASD and MSD were 95.71, 9.52, -0.84, 1.54 and 29.14, respectively. Compared with other closely related 2D-based models, the proposed method achieved the highest accuracy. In the experiment of the SLiver07, the mean values of Dice, VOE, RVD, ASD and MSD were 97.31, 5.37, -1.08, 1.85 and 27.45, respectively. Compared with other closely related models, the proposed method achieved the highest segmentation accuracy except for the RVD. Conclusion: The proposed model enables a great improvement on the accuracy compared to 2D-based models, and its robustness in circumvent challenging problems, such as small liver regions, discontinuous liver regions, and fuzzy liver boundaries, is also well demonstrated and validated.
1 Introduction
Automatic liver segmentation is clinically needed because manual CT labeling is labor-intensive and experience-dependent, while existing FCN and U-Net approaches remain insufficiently accurate and robust for difficult cases. The paper proposes SAR-U-Net, combining attention, multi-scale feature extraction, and residual learning to address these limitations.
- Motivation: Manual liver labeling from CT slices is labor-intensive and dependent on radiologist experience.Accurate liver volume, position, and shape information is needed for cancer detection and treatment.
- Related approaches: Existing automatic methods include intensity-based, machine-learning-based, and deep-learning-based approaches.Intensity methods can require human intervention, while machine-learning methods depend on manually designed features.
- Limitations of existing methods: FCN-based methods can miss pixel relationships and image details, limiting accurate liver segmentation.The passage attributes this limitation to classifying pixels without fully considering their relationships.
- Limitations of existing methods: U-Net-based methods still show insufficient accuracy and robustness when applied directly to clinical data.The paper identifies overcoming traditional U-Net shortcomings as its central concern.
- Proposed direction: SAR-U-Net introduces encoder attention, ASPP-based multi-scale extraction, and residual blocks into a U-Net framework.The proposed design aims to adaptively emphasize task-relevant features, extract richer multi-scale information, and support deeper training.
2 Related Works
The related work reviews attention mechanisms, atrous convolution, and residual learning as techniques for improving segmentation networks. SAR-U-Net combines these three techniques for automatic liver segmentation in CT.
- Attention mechanism: Attention-based networks can focus on task-specific features while suppressing irrelevant areas.The reviewed attention mechanisms include hard and soft attention according to differentiability.
- Atrous convolution: Atrous convolution expands the receptive field and helps preserve spatial resolution after down-sampling.ASPP extends dilated convolution to capture richer contextual information.
- Residual learning: Residual structures address degradation as network depth increases and have been applied to liver and tumor segmentation.Prior work embedded residual learning into U-Net-based segmentation models.
- Combined approach: SAR-U-Net integrates attention, atrous convolution, and residual learning for automatic liver segmentation in CT.The paper presents this combination as its proposed method.
3 Method
SAR-U-Net uses a symmetric encoder–decoder with residual blocks, SE attention, ASPP modules, and weighted cross-entropy training for liver CT segmentation. Its design extracts task-relevant multi-scale features while addressing class imbalance and evaluating overlap, volume, and surface accuracy.
- Architecture: SAR-U-Net uses a symmetric encoder–decoder with eight residual blocks, four pooling layers, four SE blocks, and two ASPP modules.The network takes 512×512×1 input images; the encoder extracts features and the decoder performs feature positioning.
- Residual blocks: Residual units with batch normalization and ReLU support deeper training while reducing degradation and gradient-vanishing problems.Batch normalization also reduces sensitivity to initialization and provides a regularization effect.
- SE attention: SE blocks after residual units adaptively reweight channels so the network emphasizes features relevant to liver segmentation.The SE operation compresses each channel spatially, computes channel weights, and reapplies them to the features.
- ASPP: ASPP is used in the transition and decoder output layers to capture contextual information at multiple scales.Its multi-scale semantic features help compensate for resolution loss caused by repeated down-sampling.
- Loss function: Weighted cross-entropy addresses foreground–background imbalance by assigning smaller weights to classes with more pixels.The weighting is intended to improve segmentation of relatively small liver regions.
- Evaluation metrics: Five metrics evaluate segmentation performance: Dice, VOE, RVD, ASD, and MSD.These metrics measure similarity, volume difference, relative over- or under-segmentation, average surface distance, and maximum surface distance.
4. Experiment
Experiments used LiTS17 and SLiver07 CT datasets with standardized preprocessing and augmentation to support model training and evaluation.
- Datasets: 121 of 131 LiTS17 training datasets were used for training, while 10 were reserved for testing.
- Datasets: All 20 SLiver07 training datasets were used for testing, with 512×512 CT images and 1–3 mm inter-slice spacing.
- Preprocessing: CT preprocessing transformed Hounsfield intensities to [-200, 200], applied histogram equalization, resampled z-axis spacing to 1 mm, and normalized intensities to [0,1].
- Augmentation: Data augmentation included scaling, rotation, flipping, and B-spline elastic deformation to expand training data and reduce overfitting.
4.4 Test on LiTS17-Training dataset
On LiTS17, SAR-U-Net outperformed the compared 2D models across quantitative metrics and showed robust behavior on difficult liver cases.
- Quantitative analysis: SAR-U-Net achieved the highest mean Dice of 95.71 among the four compared models.
- Quantitative analysis: SAR-U-Net improved all five evaluation metrics with small variance compared with FCN, U-Net, and Attention U-Net.
- Training behavior: SAR-U-Net produced smoother loss and accuracy curves, with training becoming stable at approximately epoch 20.
- Quantitative analysis: The performance difference between the labeled results and SAR-U-Net was statistically significant for the evaluated accuracy comparisons (p < 0.05).
- Visual segmentation results: SAR-U-Net avoided several over- and under-segmentation errors in small, discontinuous, and blurred-boundary liver cases.
4.5 Ablation analysis on LiTS17-Training datasets
Ablation experiments evaluated residual, SE, and ASPP components, showing that their combined use improved the U-Net-based model’s performance and training behavior.
- Ablation design: Five ablation experiments modified a U-Net baseline to test the contributions of residual, attention, and ASPP modules.
- Ablation results: SAR-U-Net achieved the lowest loss and highest accuracy in the ablation curves as its residual, attention, and ASPP modules increased network depth and feature extraction.
4.6 Test on SLiver07-Training dataset
On SLiver07, SAR-U-Net showed the strongest Dice performance among the compared models, faster and smoother training behavior, and robustness on difficult cases.
- Training behavior: The proposed model had the smoothest loss curve, fastest convergence, and highest accuracy in the SLiver07 training and validation analyses.
- Visual segmentation results: SAR-U-Net showed high robustness for small liver regions, discontinuous liver regions, and blurred liver boundaries.
4.7 Impact of weight factor on the loss function
The experiments compare BCE with WBCE on LiTS17 and SLiver07, showing that the weighted loss trains more smoothly and produces more accurate segmentation.
- Training and validation loss and accuracy were compared for both loss functions on LiTS17 and SLiver07.
- WBCE converges faster and more stably than BCE on LiTS17 and SLiver07.Its smoother loss curves are reported to mitigate gradient vanishing or explosion during initial training.
- BCE accuracy varies substantially early in training, whereas WBCE remains smoother and more stable.
- WBCE produces more accurate segmentation results than BCE in representative cases.The weight factor directs attention toward liver segmentation and helps avoid over-segmentation and under-segmentation.
4.8 Comparison of running time of different models
SAR-U-Net has the highest training and testing time among the compared models, reflecting its deeper architecture and added SE, ASPP, and residual modules. The paper presents this accuracy–time trade-off as meaningful for computer-aided diagnosis, while noting remaining boundary errors in lesion cases.
- SAR-U-Net has the highest training and testing time among the compared models on LiTS17 and SLiver07.The increased time is attributed to introducing residual modules, SE blocks, and ASPP.
- SAR-U-Net training takes longer, while its testing time is only slightly higher than that of other models.
- The paper treats increased computation as a meaningful trade-off because deeper feature extraction accompanies improved accuracy.
- Boundary lesions can still produce large over-segmentation or under-segmentation errors.The paper visualizes these errors in both 2D and 3D around abnormal liver boundaries.
5 Conclusion
SAR-U-Net combines SE attention, ASPP-based multi-scale extraction, residual learning, and WBCE for automatic liver CT segmentation. It achieves high accuracy and robustness across LiTS17 and SLiver07, including challenging liver regions and boundaries.
- SAR-U-Net adds SE modules to U-Net encoder convolution units to emphasize relevant segmentation features and suppress irrelevant areas.
- ASPP serves as the transition and output layers to enlarge the receptive field and extract multi-scale features.
- Residual structures enable a deeper network while helping avoid gradient vanishing, and WBCE provides more targeted training.
- SAR-U-Net shows high accuracy and robustness on LiTS17 and SLiver07, including small, discontinuous, and blurry liver regions.The paper also reports smoother accuracy and loss curves, faster convergence, and improved ablation metrics.