4. 2. Related Works
• COSMOS
• Multiple orientation data compensate missing values
• Gold standard algorithm
• Long acquisition time
Orientation 1 Orientation 2 Orientation 3 COSMOS
5. 2. Related Works
• Conventional algorithms
• Filling conical surface, iterative algorithms
• Streaking artifact
• High computational complexity
• Deep learning algorithms
• Improved performance, fast reconstruction
• Supervised learning Ground truth
• Training data ≠ test data Performance↓
7. 3. Methods
• Training & test data
• Resolution of training data ≠ resolution of test data
• Resolution-agnostic QSM reconstruction method is required
Training Test
11. 3. Methods
• Training & test data
• Resizing by interpolation synthetic multi-resolution phase data
• In-plane: 0.424~1.28, Axial: 0.427~2.13
• Resolution range of the resized training data covers the resolution of the test data
Training Test
12. 3. Methods
• Comparison methods
• Conventional algorithms: TKD, MEDI, iLSQR
• Fidelity imposed network edit (FINE)
- Supervised learning fine tuning with test data using unsupervised fidelity loss
• Meta-QSM
- Resolution-arbitrary network for QSM reconstruction
- Weight-predict convolutional layers
13. 4. Experimental Results
• Effects of AdaIN transform / unsupervised learning
• AdaIN IN
• Artifacts (yellow arrows)
• Unsupervised supervised
• Cannot reconstruct detailed structures
• Supervised learning is more sensitive to lack
of training data than unsupervised learning
19. 5. Discussion & Conclusion
• Conventional methods
• Noticeable noise, artifacts (e.g. streaking artifact, digitized artifact)
• Time consuming
• Deep learning algorithms
• Under/overestimation
• Blurry output
• Fail to restore susceptibility maps of some resolution data
• Supervised learning sensitive to lack of training data
• Proposed method
- Reconstruct QSM of various resolutions with less artifacts and noise
- Unsupervised learning does not suffer from problems of the supervised learning methods