Optimizing thoracic synthetic computed tomography generation from magnetic resonance imaging: the role of Fourier transform and other key factors.

BACKGROUND AND PURPOSE: Magnetic Resonance Imaging-only (MRI-only) workflows are an emerging strategy in radiotherapy, with artificial intelligence (AI) playing a central role in generating synthetic computed tomography (sCT) images. The thorax remains a particularly difficult region due to marked electron density (ED) heterogeneity and respiratory motion. This study investigates the impact of key factors on AI-based thoracic sCT generation.

MATERIALS AND

METHODS

A total of 122 thoracic patients treated with MRI-guided radiotherapy (MRIgRT) were retrospectively included. Both 0.35 Tesla (T) MR and CT simulation images were acquired under consistent breath-hold conditions. Three aspects were analyzed: (i) training set size (34, 68, and 102 cases), (ii) pre-processing of MR images (filtered versus raw), and (iii) generator architecture, comparing U-Net and ResNet with a novel model integrating Fourier space information, the Adaptive Fourier Neural Operator (AFNO). Models were tested on 20 independent patients using image similarity metrics. The best configuration was also evaluated through dose recalculations.

RESULTS

Expanding the training set improved accuracy, reducing Mean Absolute Error (MAE) from 42.0 ± 9 Hounsfield Units (HU) to 35.9 ± 6 HU. Pre-processing had limited effect, while generator architecture had a strong impact, with AFNO outperforming others (MAE = 32.4 ± 6 HU). The optimal setup, AFNO trained on raw MR images from 102 patients, yielded dosimetric deviations below 3 % for target dose-volume metrics and within 50 cGy for organs at risk (OARs).

CONCLUSIONS

These findings highlight the importance of training dataset size and advanced network architectures for thoracic sCT generation. AFNO demonstrated superior performance, reinforcing the feasibility of MRI-only workflows in thoracic radiotherapy.