Characterize registration-induced bias in supervised synthetic CT evaluation

Characterize how structured or systematic residual registration errors affect voxel-wise regression, anatomical fidelity, benchmark rankings, and the interpretation of reference-based metrics in supervised synthetic CT generation.

Background

Supervised synthetic CT methods are commonly evaluated with intensity-based metrics computed against registered CT references. Because residual multimodal misregistration can alter anatomical structures and intensity correspondences, these metrics may reward adaptation to a registration convention rather than preservation of patient-specific anatomy.

The paper identifies this limitation as insufficiently characterized in the literature. Resolving it requires determining how registration-induced geometric discrepancies influence training behavior, metric sensitivity, benchmark rankings, and the relationship between voxel-wise agreement and anatomical fidelity.

References

This limitation remains insufficiently characterized in the recent literature , where evaluation protocols predominantly emphasize intensity-based similarity rather than anatomical faithfulness.

— When Misalignment Becomes Supervision: Structured Label Noise in Supervised Synthetic CT Generation  (2609.29387 - Boussot et al., 24 Sep 2026) in Section 2, paragraph “Consequences for synthetic CT evaluation”

Errors in these earlier estimates could therefore propagate into subsequent registrations, and the robustness of the memory mechanism to such error propagation remains to be evaluated.

— Personalising Deformable Image Registration with Clinical Context and Deformation History for Adaptive Radiotherapy  (2610.18926 - Jiang et al., 13 Oct 2026) in Discussion, paragraph beginning “These findings should nevertheless be interpreted in light of several limitations.”