Laser & Optoelectronics Progress, Volume. 60, Issue 12, 1210001(2023)
Fusion of Attention Mechanism and Deformable Residual Convolution for Liver Tumor Segmentation
Figures & Tables(13)
Fig. 4. Schematic of deformable convolution. (a) Traditional convolution kernel; (b) deformable convolution kernel
Fig. 8. Data pre-processing. (a) Original CT image; (b) segmentation result of ribs and spine; (c) cropping diagram; (d) pre-processed image
Table 1. Comparison results of ablation experiments
View tableTable 1. Comparison results of ablation experiments
Module Dice /% VOE /% RVD /% ASD /mm MSSD /mm U-Net 71.8 48.5 -25.3 2.44 9.13 U-Net+new skip-connection 73.1 47.2 -22.1 2.01 9.01 U-Net+dual attention 79.3 37.2 -11.5 1.69 8.13 U-Net+deformable Conv 76.8 40.1 -12.4 1.56 7.92 U-Net+RBE 82.0 36.7 -7.3 1.34 7.32 U-Net+ deformable Conv + dual attention 81.9 38.8 -8.7 1.42 7.54 U-Net+RBE+deformable Conv+dual attention 83.8 35.3 -6.2 1.29 7.21 U-Net+all modules 85.2 32.4 -3.2 0.98 6.61
Table 2. Performance comparison of different methods on LITS test set
View tableTable 2. Performance comparison of different methods on LITS test set
Method Dice /% VOE /% RVD /% ASD /mm MSSD /mm U-Net［ 14 ］71.8 48.5 -25.3 2.42 9.13 BS-Unet［ 15 ］75.1 44.0 -23.4 1.73 8.54 A-Unet［ 17 ］76.2 40.6 -14.4 1.42 7.92 U-Net++［ 18 ］80.8 37.2 -10.7 1.26 7.12 U-Net3+［ 20 ］81.2 36.9 -11.2 1.31 7.41 MDCC-Unet［ 26 ］83.9 34.1 -6.6 1.22 7.53 proposed method 85.2 32.2 -3.2 0.98 6.61
Table 3. Performance comparison with other methods on LITS dataset
View tableTable 3. Performance comparison with other methods on LITS dataset
Method Dice /% VOE /% RVD /% ASD /mm MSSD /mm RA-UNet［ 27 ］59.5 38.9 -15.2 1.289 6.775 Method in Ref.［ 28 ］67.0 45.0 4.0 6.660 57.930 Method in Ref.［ 22 ］67.6 34.1 -6.4 — — Method in Ref.［ 29 ］73.69 37.80 -15.78 — — MA-Net［ 30 ］74.9 21.0 -18.0 — — CUResNets［ 31 ］75.2 37.9 -15.9 — — Method in Ref.［ 32 ］83.32 11.62 -15.98 — — Proposed method 85.2 32.2 -3.2 0.98 6.61
Tools
Get Citation
Copy Citation Text
Wenhan Yang, Miao Liao. Fusion of Attention Mechanism and Deformable Residual Convolution for Liver Tumor Segmentation[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1210001
Paper Information
Category: Image Processing
Received: Apr. 20, 2022
Accepted: May. 25, 2022
Published Online: May. 23, 2023
The Author Email: Miao Liao (Liaomiaohi@126.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20