Get excited about our newest development from Pymedix R&D!

Kris T. Huang, MD, PhD, CTO

Mapping image registration confidence

Automated TG-132 patient-specific QA

Pre- (green) and post-treatment (magenta) CT fusion

Registration confidence heat map
(Red: low, blue/green: high)

We’re excited to announce our newest development from Pymedix R&D: the registration confidence map! We gave a sneak peek of the foundation of registration confidence assessment in our peer-reviewed journal article, and we’ve taken it to its logical conclusion – a full spatial map based on patient-specific information, not aggregate or inferred data. And yes, we can take more traditional metrics like the Jacobian and shear deformation into account.

Note that this is a confidence map which estimates potential error. Homogeneous regions, like the brain (on CT) or the liver (on pretty much any modality) intrinsically don’t contain much, if any, information to register. That doesn’t mean Autofuse’s registration is wrong; it’s just an honest portrayal of how much error there could be, which is just as important to know as how right it (likely) is.

For fast, easy visualization, Our color-coding conveniently corresponds roughly to TG-132’s registration uncertainty assessment levels:

Autofuse confidence map color	TG-132 registration uncertainty assessment level	TG-132 description
Blue/green	1	Anatomy local to the area of interest is undistorted and aligned within 1 mm Useful for structure definition within the local region Appropriate for localization provided target is in locally aligned region
Yellow	2	Aligned locally, with mild anatomical variation Acceptable registration required deformation which risks altering anatomy Registered image shouldn’t be used solely for target definition as target may be deformed Increased reliance on additional information is highly recommended Registered image information should be used in complementary manner and no image should be used by itself
Red	3	Registration not good enough to rely on geometric integrity Possible use to identify general location of lesion (e.g., PET hot spot)
Red	4	Unable to align anatomy to acceptable levels Patient position variation too great between scans (e.g., surgical resection of the anatomy of interest or dramatic weight change between scans)

A full spatial map is crucial for automated patient-specific QA. In particular, it allows for:

Global registration assessment for the entire image
Regional assessment
Registration assessment per segmented structure, which in turn enables:
- Automated TG-132 reports!
- Daily automated adaptive assessment

And yes, we use perceptually uniform colormaps for all data displays. (Here’s why this matters!)