ESTRO 2024 - Abstract Book

S4117

Physics - Inter-fraction motion management and offline adaptive radiotherapy

ESTRO 2024

In the real data experiment, we scanned a head-and-neck phantom with an image size being 512x512x93. We first acquired full-scan projections at a reference position and then acquired limited-angle projections with 80 o discrepancy at a moving position where the couch was moved with both translations and rotations along the 3 axes. We evaluated the performance of the real data registration results by comparing the acquired projections with the projection generated from the registered 3D volume using the learned registration parameters. The Peak Signal-to Noise Ratio (PSNR) of between the acquired and generated projections converged to 21.23 dB within 10 minutes.

Conclusion:

We propose a dataset-free and lightweight MLP for the challenging 2D-3D registration problem. No dataset but only two projections with limited-angle discrepancy at the new position were sufficient for the registration. The results demonstrated the accuracy and efficiency of our method in both simulation and real data experiments.

Keywords: 2D-3D registration, CBCT, neural representation

3072

Digital Poster