Intuitively, the task of registering two images (for example, aligning a preoperative CT image with an intraoperative radiograph or cone-beam CT) must depend on the quality of the images. And it stands to reason that the accuracy of registration will improve with the quality of those images. But what is the connection – exactly – and what are the image quality factors that govern registration accuracy? Spatial resolution? Noise? And are the limits in visual image quality (for example, a low-dose image for which a feature is no longer visible) the same as the lower limits in registration performance?
These questions are at the heart of a new paper by Michael Ketcha and co-authors at the I-STAR Lab in Biomedical Engineering at Johns Hopkins University, yielding a theoretical model that links image registration performance with image quality. Models for each have been established in previous work, but the connection between the two has not been well formulated. For example, Michael Fitzpatrick and colleagues established a statistical framework for understanding Target Registration Error (TRE), governed by the Fiducial Localization Error (FLE), Fiducial Registration Error (FRE), and the spatial distribution of fiducials with respect to a target point. Meanwhile, Ian Cunningham and colleagues produced a cascaded systems model for image quality describing the propagation of signal and noise – providing the basis for image quality models describing the tradeoffs among spatial resolution, noise, and dose in flat-panel x-ray detectors, tomosynthesis, and cone-beam CT. Such theoretical models have been invaluable to the development of new imaging and image guidance systems over the last two decades, but the connection between the two – HOW IMAGE QUALITY AFFECTS REGISTRATION ACCURACY – has remained largely unanswered.
Michael Ketcha’s paper published in IEEE-TMI in July 2017 derives the Cramer-Rao lower bound (CRLB) for registration accuracy in a manner that reveals the underlying dependencies on spatial resolution and image noise. By analyzing the CRLB as a function of dose, the work sheds light on the low-dose limits of image registration in a manner that could help reduce dose in image-guided interventions, where the task is often one of registration rather than visual detection.
The analysis considers the CRLB as the inverse of the Fisher Information Matrix (FIM) and derives the relationship on two main factors. First is the image noise, which depends on dose and may be different in the two images. Second is the power of (sum of squared) image gradients, which is governed by the contrast and frequency content of the subject. The FIM is thereby related to factors of image noise, resolution, and dose in a manner that permits analysis of the CRLB for a variety of scenarios – including registration of low-contrast soft tissues, high-contrast bone structures, and the effect of image smoothing to improve registration performance.
The work is analogous to widespread efforts to identify low-dose limits of visual detectability via models of imaging task. In image-guided interventions, however, the task of registration is often as important (or more important) as the task of visualization, allowing preoperative images and planning information to be accurately aligned with the patient at the time of treatment. Previous experiments by Uneri et al. showed that registration algorithms can perform well at dose levels below that which would normally be considered to yield a visually acceptable image — effects that are borne out by Ketcha’s analysis.