Towards Fully Automatic X-Ray to CT Registration

Abstract

The main challenge preventing a fully-automatic X-ray to CT registration is an initialization scheme that brings the X-ray pose within the capture range of existing intensity-based registration methods. By providing such an automatic initialization, the present study introduces the first end-to-end fully-automatic registration framework. A network is first trained once on artificial X-rays to extract 2D landmarks resulting from the projection of CT-labels. A patient-specific refinement scheme is then carried out: candidate points detected from a new set of artificial X-rays are back-projected onto the patient CT and merged into a refined meaningful set of landmarks used for network re-training. This network-landmarks combination is finally exploited for intraoperative pose-initialization with a runtime of 102 ms. Evaluated on 6 pelvis anatomies (486 images in total), the mean Target Registration Error was 15.0 ± 7.3 mm. When used to initialize the BOBYQA optimizer with normalized cross-correlation, the average (± STD) projection distance was 3.4 ± 2.3 mm, and the registration success rate (projection distance < 2.5 % of the detector width) greater than 97 %.