Summary
Generating synthetic CT scans from MRI or cone beam CT images often relies on comparing each voxel to a matched real CT scan. However, the computer alignment between these images isn’t perfect and causes consistent geometric errors rather than random noise. The authors studied thousands of patient image pairs to show that this misalignment biases training and evaluation, with models partly learning the registration errors. They found that using more anatomically consistent alignment reduces variability and improves robustness. To address this, they proposed a new way to train the models using a perceptual loss that focuses on anatomical features, leading to sharper and more realistic synthetic CT images.
What this means in practice
- •For medical imaging developers: Improve synthetic CT generation tools by incorporating anatomy-aware training losses that reduce the impact of registration errors.
- •For clinical radiology teams: Enhance the reliability of synthetic CT scans used in treatment planning by adopting evaluation metrics sensitive to anatomical accuracy rather than only voxel matches.
Authors
Valentin Boussot, Cedric Hemon, Caroline Lafond, Jean-Claude Nunes, Jean-Louis Dillenseger
Abstract
Supervised synthetic CT (sCT) generation is commonly trained and evaluated as voxel-wise regression against registered reference CT images. In practice, MRI-CT and CBCT-CT pairs are aligned through registration procedures that leave residual misalignments. These residuals are not independent intensity noise but spatially coherent geometric discrepancies that act as structured label noise. We investigate how this registration-induced bias affects supervised MRI-to-CT and CBCT-to-CT synthesis on 1,784 paired patients covering five anatomical regions. Voxel-wise scores strongly depend on the consistency between the registration used to build the training targets and the one used for evaluation: models score best when both conventions match, showing that networks partly learn the geometric convention of the registration pipeline and that standard metrics reward it. Training on more anatomically consistent registrations reduces prediction variability and improves out-of-distribution robustness, and CT-only controls show that registration alone produces metric errors in the range of top challenge submissions. To mitigate the limits of voxel-wise supervision, we introduce a perceptual loss computed in the feature space of a pretrained Segment Anything encoder. Compared with MAE-only and VGG-based objectives, it improves downstream segmentation and yields sharper, more structurally coherent sCT. Perceptual and voxel-wise metrics disagree under imperfect alignment and agree when the evaluation geometry is reliable. These results identify registration-induced bias as a central confounder in supervised sCT generation and argue for complementing voxel-wise agreement with anatomy-oriented evaluation criteria.