Multi‑ Attention Network for

Automatic Liver Tumor Segmentation

Juan Pablo Betancur, Aron Gimesi,

Tuhinangshu Gangopadhyay
 Liver Tumor Segmentation
Detection Segmentation Benefits

Liver cancer was ● Accurate Diagnosis

among the top three ● Treatment Planning
causes of cancer death ● Personalized Medicine
in 46 countries [5] ● Monitoring Progression
Example of Liver Tumor Segmentation

Approaches Challenges
● Tumor Variability
● Deep Learning: U-Net, UNet++,
● Similarity Issues
Attention U-Net, Residual ● Technical Factors
Learning.
● Traditional Machine Learning:
SVM, NB, KNN.
2
 MANet-Architecture[1]
Key concepts related to this architecture:
• U-Net
• Residual Blocks
• Skip connection attention gate
• Channel Attention
– Squeeze and Excite attention
• Spatial Attention
• Convolutional block attention module (CBAM)

3
 U-Net + Residual Blocks
Residual
U-Net
Learning

Ronneberger et al. (2015), "U-Net: Convolutional Networks for Biomedical Image He et al. (2016), "Deep Residual Learning for Image Recognition" [4]
Segmentation" [3]

Pros of
merging
● Enhanced Feature Propagation
● Preservation of Important Features
● Better handling of Degradation
● Synergy of High-Resolution and
Semantic Information
● Improved segmentation performance 4
Skip connection attention gate[6]

5
Squeeze and Excite Attention[2]

6
Channel Attention[7]

7
Spatial Attention[7]

8
Convolutional block attention module (CBAM)[7]

9
 Final MANet Architecture

Residual Learning is
The architecture
used to enhance
resembles a U-Net,
feature propagation
utilizing skip
and
connections to
mitigate vanishing
preserve and
gradient.
integrate high- and
low-level features
during upsampling.

10
 Dataset
• LiTS17 challenge dataset + 3DIRCADb
• In total:
– 130 CT scans (512 x 512 x SLICES)
– 7050 slices (512 x 512)
• Number of tumors per scan: 0 – 75
• Size of the tumors: 38 mm3 – 349 cm3
• No validation set (not enough data)
• 4:1 training-test set ratio

11
 Preprocessing
• Truncating intensities to enhance liver visibility
(-150 – 250 HU)
• Histogram equalization
• Normalization (0 – 1)
• Data augmentation
– flipping
– shifting
– scaling
– rotation

12
 Results

• MANet outperforms baseline models in most metrics, while

maintaining similar or even less parameter count compared
to some baselines (7.8M)

13
 Results (Quantitative)

Task Method Dice score ASSD IoU VOE (1-IOU) Accuracy Sensitivity Specificity

Volume-based UNet 0.6612 ± 0.277 1.0843 ± 1.425 0.5469 ± 0.266 0.4530 ± 0.266 0.9950 ± 0.004 0.6394 ± 0.285 0.9987 ± 0.002

Attention UNet 0.6505 ± 0.278 1.2551 ± 1.338 0.5356 ± 0.263 0.4643 ± 0.263 0.9945 ± 0.006 0.6250 ± 0.292 0.9984 ± 0.002

UNet + Resnet18 0.6560 ± 0.281 0.9321 ± 0.960 0.5433 ± 0.268 0.4566 ± 0.268 0.9950 ± 0.005 0.6108 ± 0.294 0.9991 ± 0.001

UNet + CBAM 0.6635 ± 0.271 1.2795 ± 1.638 0.5487 ± 0.261 0.4512 ± 0.261 0.9946 ± 0.005 0.6678 ± 0.283 0.9981 ± 0.002

MANet 0.6735 ± 0.267 1.2049 ± 1.356 0.5590 ± 0.258 0.4409 ± 0.258 0.9950 ± 0.004 0.7426 ± 0.283 0.9978 ± 0.002
(Proposed)
Slice-based UNet 0.7790 ± 0.208 0.9009 ± 1.020 0.6744 ± 0.217 0.3255 ± 0.217 0.9940 ± 0.006 0.7476 ± 0.237 0.9982 ± 0.001

Attention UNet 0.7676 ± 0.195 0.9188 ± 0.783 0.6550 ± 0.208 0.3449 ± 0.208 0.9935 ± 0.006 0.7423 ± 0.231 0.9978 ± 0.002

UNet + Resnet18 0.7686 ± 0.211 1.0037 ± 1.429 0.6619 ± 0.223 0.3380 ± 0.223 0.9934 ± 0.007 0.7342 ± 0.245 0.9984 ± 0.001

UNet + CBAM 0.7784 ± 0.202 0.8241 ± 0.870 0.6784 ± 0.224 0.3279 ± 0.224 0.9941 ± 0.004 0.7439 ± 0.245 0.9982 ± 0.002

MANet 0.8145 ± 0.150 0.7084 ± 0.701 0.7084 ± 0.171 0.2915 ± 0.171 0.9947 ± 0.004 0.8723 ± 0.173 0.9970 ± 0.002
(Proposed)

14
Results (Qualitative)

15
Results (Qualitative)

16
 Conclusion
• MANet: A novel multi-attention network for liver tumor
segmentation
• Key features:
– U-Net based architecture with residual blocks
– Multiple attention mechanisms: SCAG, CA, SA, CBAM
• Advantages:
– Improved feature selection and focus on relevant areas
– Better performance on challenging cases (small, low-contrast
tumors)
– Competitive results with relatively small parameter overhead
• Future work:
– Further optimization for clinical application
– Improvement of volume-based implementation

17
 References
[1] Hettihewa, Kasun, et al. "MANet: a multi-attention network for automatic liver tumor
segmentation in computed tomography (CT) imaging." Scientific Reports 13.1 (2023):
20098.
[2] Hu, Jie, Li Shen, and Gang Sun. "Squeeze-and-excitation networks." Proceedings of the IEEE
conference on computer vision and pattern recognition. 2018.
[3] Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-Net: Convolutional networks for
biomedical image segmentation." International Conference on Medical Image Computing
and Computer-Assisted Intervention. Springer, Cham, 2015.
[4] He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Deep residual learning for image
recognition." Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition (CVPR). 2016: 770-778.
[5] Ferlay, Jacques, et al. "Global burden of primary liver cancer in 2020 and predictions to
2040." Journal of Hepatology 77.6 (2022): 1598-1606.
DOI: 10.1016/j.jhep.2022.08.021
[6] Vaswani, A. "Attention is all you need." Advances in Neural Information Processing Systems
(2017).
[7] Woo, Sanghyun, et al. "Cbam: Convolutional block attention module." Proceedings of the
European conference on computer vision (ECCV). 2018.

18
 Read more!

• U-Net: https://doi.org/10.1007/978-3-319-24574-4_28
• Residual Networks:
https://openaccess.thecvf.com/content_cvpr_2016/papers/He_Deep_Residua
l_Learning_CVPR_2016_paper.pdf
• Skip connection attention gate:
• https://user.phil.hhu.de/~cwurm/wp-content/uploads/2020/01/7181-attentio
n-is-all-you-need.pdf
• Squeeze and Excite Attention:
https://openaccess.thecvf.com/content_cvpr_2018/html/Hu_Squeeze-and-Ex
citation_Networks_CVPR_2018_paper.html
• Channel Attention, Spatial Attention, CBAM:
https://openaccess.thecvf.com/content_ECCV_2018/html/Sanghyun_Woo_Co
nvolutional_Block_Attention_ECCV_2018_paper.html

19
THANK YOU! :)

20