Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
ST: Fachverband Strahlen- und Medizinphysik
ST 1: Computational Methods and Simulation
ST 1.6: Vortrag
Dienstag, 1. April 2025, 15:00–15:15, ZHG003
Cycle GAN-Based Style Transfer for Image Registration between Clinical and HiP-CT — Lukas Johanns1, Michael Windau1, Lucas Cremer1, •Claire Walsh2, Joe Jacob2, Joseph Brunet3, Paul Sweeney4, and Stijn Verleden5 — 1TU Dortmund — 2UCL London — 3ESRF Grenoble — 4Cancer Research UK — 5UZA Antwerp
Registration is an image processing algorithm that enables the alignment of scans across domains. It finds applications in disease tracking and research, in the operating room, and as a preprocessing step for other machine learning algorithms.
However, registration across modalities requires careful tuning and preprocessing of the dataset, as different domains are often difficult to compare.
To improve the performance of registration algorithms, a Cycle-GAN model for style transfer is used as a preprocessing step. This model transfers the style of the target domain to the input image to enhance registration performance. This project investigates the performance and feasibility of such a deep learning model applied to Clinical-CT scans and high-resolution/high-contrast HiP-CT scans (Hierarchical Phase Contrast Tomography). In this talk, the concepts of registration and Cycle-GANs are briefly introduced. Afterwards, the results of our style-transfer network are presented, followed by a discussion of a future feasibility study for registration.
Keywords: Image Registration; Style Transfer; HiP-CT; Deep Learning; Cycle-GAN