A Multi-Filter and Multi-Scale U-Net for Cone-Beam Computed Tomography with Hardware Constraints

Andreas Hauptmann, Mustafa Al-Rubaye, Miika T. Nieminen, Mikael A. K. Brix

Abstract: Learned reconstructions for 3D cone-beam computed tomography (CBCT) require significant hardware resources for training as well as evaluation. In this challenge paper we aim to improve performance of the U-Net architecture for post-processing by creating multiple inputs to the network using varying frequency filters. The networks are able to be trained on a single GPU and achieved 3rd place in the ICASSP 2024 3D-CBCT grand challenge.