UCLA

Deformable image registration (DIR) is central to thoracic radiation therapy, relating anatomy across the breathing cycle and the treatment course so that dose can be tracked and adapted. Its value depends on each registration being correct, yet verifying correctness remains unsolved. Similarity scores can report excellent agreement for registrations that are demonstrably wrong, and target registration error (TRE), the accepted reference standard, is evaluated at a handful of expert-placed landmarks that characterize accuracy only where an observer identified a landmark. A dense, automated measure of registration error is needed to verify DIR with confidence. GPU-accelerated software was developed to extract anatomical landmarks from thoracic CT, place them in correspondence across a breathing-resolved image ensemble, measure registration error densely from that correspondence, and test an interpretable similarity score against the result. Each stage was optimized for the architecture of graphics processing units (GPUs) for near-interactive performance. A medial-axis extraction was developed to drive candidate voxels onto vessel centerlines directly from the CT intensity field by iterative mean shift, without a prior segmentation or user-driven point selection. The resulting point cloud was assembled into a skeleton graph, its breaks repaired by a directional signature whose branch directions are selected by topological persistence. A cross-scan correspondence method identified the same anatomical landmarks across a 25-image 5DCT ensemble, certified by majority vote and measured by anatomical consistency. These landmarks provide a registration-independent ground truth, against which the TRE of a registration under test is measured at every landmark, yielding a spatially resolved error field rather than a single value. An interpretable similarity score, λ, was developed by adapting the γ dose-comparison metric to CT intensity. Validated against the dense error field, λ tracked registration error only weakly, no better than standard metrics such as NCC and SSIM, revealing that its magnitude follows image-gradient availability rather than displacement. This establishes that a physically constructed, interpretable score cannot substitute for a geometric reference, and argues for the dense reference itself. The methods were demonstrated on a 5DCT ensemble but developed as general techniques, intended to extend across patients and scanning protocols. Together they provide a foundation for verifying thoracic CT deformable image registration in an independent, rapid, and automated fashion.