Accurate registration of medical images—especially across time—is essential for cancer treatment planning, organ tracking, and surgical guidance. Yet, traditional deformable image registration (DIR) often fails when faced with large deformations, noise, or missing data.
This project investigates an advanced approach using Implicit Neural Representations (INRs). Instead of grids or control points, INRs encode deformation fields as neural networks, enabling smooth, high-resolution, and data-efficient registration.
You'll help implement and evaluate an INR-based DIR method and compare it with classic techniques in 3D image registration tasks.
Objectives of the work (specific tasks will be agreed individually with the student)
• Understand basics of deformable image registration and implicit neural networks.
• Implement an INR-based model to encode 3D deformation fields.
• Apply the model to synthetic and real medical imaging datasets (e.g. CT, MRI).
• Benchmark the method against classical DIR (e.g. B-spline, Demons, SDF-based).
Benefits for the student
• Learn cutting-edge neural techniques for scientific computing.
• Gain experience with 3D medical image analysis and preprocessing.
• Possibility to co-author publications in medical imaging and AI.
• Have a possible chance to work with Czech–Korean researchers on a real-world health challenge.
Resources
- Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., ... & Sridhar, S. (2022, May). Neural fields in visual computing and beyond. In Computer Graphics Forum (Vol. 41, No. 2, pp. 641-676).
- Sitzmann, V., Martel, J., Bergman, A., Lindell, D., & Wetzstein, G. (2020). Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33, 7462-7473.
- Modersitzki, J. (2003). Numerical methods for image registration. OUP Oxford.
