Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 8, 1090(2024)
Image registration combining cross-scale point matching and multi-scale feature fusion
Fig. 2. Structure diagram of cross-scale point matching combined with multi-scale feature fusion network
Fig. 4. Structure diagram of the multi-scale feature fusion module（a）and SCSE module（b）
Fig. 6. Comparison of results from 5 registration methods on two datasets
Fig. 7. Visualization results of ablation experiments on two datasets
Fig. 8. Box plot of Dice values for 27 anatomical structures in the OASIS-3 dataset
Fig. 9. Bar chart of average Dice values for 27 anatomical structures in the ABIDE dataset
Table 1. Comparison of 5 registration methods on two different datasets（average±standard deviation）
View tableView in ArticleTable 1. Comparison of 5 registration methods on two different datasets（average±standard deviation）
Method OASIS-3 LPBA40 Dice ASD Dice ASD Affine 0.481±0.021 1.870±0.110 0.594±0.012 1.620±0.146 SyN 0.592±0.014 1.230±0.085 0.664±0.011 1.220±0.098 VoxelMorph 0.707±0.012 0.839±0.062 0.640±0.028 1.316±0.126 CycleMorph 0.695±0.012 0.872±0.060 0.629±0.027 1.335±0.127 CM-RegNet 0.716±0.012 0.796±0.064 0.653±0.027 1.270±0.113
Table 2. Comparison of the average registration time of 5 registration methods on two different datasets s
View tableView in ArticleTable 2. Comparison of the average registration time of 5 registration methods on two different datasets s
Method OASIS-3 LPBA40 Affine 5.39 3.58 SyN 23.50 15.69 VoxelMorph 0.15 0.14 CycleMorph 0.16 0.14 CM-RegNet 0.61 0.43
Table 3. Comparison of ablation experiment results of CM-RegNet model（average±standard deviation）
View tableView in ArticleTable 3. Comparison of ablation experiment results of CM-RegNet model（average±standard deviation）
Baseline CPM MFF SCSE OASIS-3 LPBA40 √ 0.707±0.012 0.640±0.028 √ √ 0.709±0.012 0.651±0.026 √ √ 0.709±0.016 0.648±0.019 √ √ √ 0.713±0.011 0.650±0.021 √ √ √ √ 0.716±0.012 0.653±0.027
Table 4. Comparison of accuracy of several registration methods in five anatomical structures
View tableView in ArticleTable 4. Comparison of accuracy of several registration methods in five anatomical structures
Method Accuracy/% White Matter Inf. Lat. Vent. Brain Stem CSF Vessel Affine 59.8 14.9 77.6 40.5 9.6 SyN 65.7 13.8 85.6 51.0 2.9 VoxelMorph 83.1 47.4 91.7 69.3 28.4 CycleMorph 82.5 43.9 90.9 68.6 25.2 CM-RegNet 84.2 49.2 92.0 70.3 30.8
Table 5. Comparison of generalization results of several methods on ABIDE dataset（average±standard deviation）
View tableView in ArticleTable 5. Comparison of generalization results of several methods on ABIDE dataset（average±standard deviation）
Method Dice ASD Affine 0.462±0.036 1.698±0.121 SyN 0.570±0.020 1.054±0.074 VoxelMorph 0.564±0.126 1.369±0.444 CycleMorph 0.546±0.121 1.396±0.398 CM-RegNet 0.574±0.132 1.301±0.369
Get Citation
Copy Citation Text
Zhuolin OU, Xiaoqi LÜ, Yu GU. Image registration combining cross-scale point matching and multi-scale feature fusion[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(8): 1090
Category: Object Tracking and Recognition
Received: Aug. 24, 2023
Accepted: --
Published Online: Sep. 27, 2024
The Author Email: Xiaoqi LÜ (lxiaoqi@imut.edu.cn)
