Monai for Deep-Learning Based CBCT Reconstruction

As part of the ICASSP 2024 CBCT challenge, this paper presents the application of the MONAI framework to 3D cone-beam computed tomography (CBCT) reconstruction. Employing the SegResNet architecture, two networks were trained on clinical and low dose reconstructions, respectively. The models, when tested on held-out data, achieved mean squared errors of 8.41e-4 for clinical dose and 1.476e-3 for low dose reconstructions. Notably, the proposed solution secured a second-place position in the challenge.