[Paper Review] 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+2 more|arXiv (Cornell University)|Mar 11, 2021

Advanced Neural Network Applications32 references4 citations

TL;DR

This paper proposes SAR-U-Net, a 2D U-Net variant enhanced with squeeze-and-excitation (SE) blocks, atrous spatial pyramid pooling (ASPP), and residual learning for accurate liver segmentation in CT scans. The model achieves state-of-the-art performance on LiTS17 and SLiver07 datasets, with Dice scores of 95.71% and 97.31%, respectively, demonstrating superior robustness to challenging liver boundaries and small/fragmented regions.

ABSTRACT

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.

Motivation & Objective

• To address the challenges of inaccurate liver segmentation in CT scans due to small liver regions, discontinuous structures, and fuzzy boundaries.
• To enhance feature representation and multi-scale context modeling in 2D U-Net architectures for improved segmentation accuracy.
• To reduce degradation in deep networks by integrating residual learning into the U-Net encoder and decoder.
• To validate the effectiveness of combining SE blocks and ASPP in a U-Net framework for medical image segmentation.
• To achieve state-of-the-art performance on public liver segmentation benchmarks, LiTS17 and SLiver07.

Proposed method

• The SE block is integrated after each convolutional layer in the U-Net encoder to recalibrate channel-wise feature responses based on global context, emphasizing relevant features and suppressing irrelevant ones.
• Atrous Spatial Pyramid Pooling (ASPP) replaces the transition and output layers to capture multi-scale contextual information through dilated convolutions with different rates.
• Residual blocks replace standard convolutional blocks to mitigate gradient vanishing and enable training of deeper networks, improving learning efficiency and accuracy.
• The encoder-decoder architecture of U-Net is preserved, with skip connections facilitating feature fusion between contracting and expansive paths.
• The network is trained end-to-end using binary cross-entropy and dice loss to optimize segmentation performance.
• The model is evaluated on two public datasets: LiTS17 and SLiver07, using standard metrics including Dice, VOE, RVD, ASD, and MSD.

Experimental results

Research questions

• RQ1 Can the integration of squeeze-and-excitation blocks improve feature representation in a U-Net for liver CT segmentation?
• RQ2 Does incorporating atrous spatial pyramid pooling enhance multi-scale context modeling and segmentation accuracy in 2D U-Net architectures?
• RQ3 To what extent does residual learning improve training depth and performance in a U-Net-based liver segmentation model?
• RQ4 How does the proposed SAR-U-Net compare to state-of-the-art 2D models on benchmark liver segmentation datasets?
• RQ5 Can the model robustly handle challenging cases such as small, discontinuous, or poorly defined liver regions?

Key findings

• On the LiTS17 dataset, SAR-U-Net achieved a Dice score of 95.71%, outperforming other 2D-based models in the comparison.
• The model recorded a VOE of 9.52%, RVD of -0.84%, ASD of 1.54 mm, and MSD of 29.14 mm on LiTS17, indicating high segmentation accuracy and consistency.
• On the SLiver07 dataset, the model achieved a Dice score of 97.31%, the highest among comparable models, with VOE of 5.37%, RVD of -1.08%, ASD of 1.85 mm, and MSD of 27.45 mm.
• The model demonstrated strong robustness in segmenting small, discontinuous, and poorly defined liver regions, outperforming baseline U-Net and related architectures.
• The ablation study confirmed that the combination of SE blocks, ASPP, and residual learning significantly improves performance over individual components.
• The results validate that the proposed SAR-U-Net architecture effectively captures multi-scale context and enhances feature discrimination, leading to state-of-the-art segmentation accuracy.