Author: Edward Robert Criscuolo, Deshan Yang 👨🔬
Affiliation: Duke University, Department of Radiation Oncology, Duke University 🌍
Purpose:
Deformable registration of 4DCT images has many clinical applications, but current methods are unreliable and can produce dangerous errors. Iterative, parametrized image registration does perform highly on existing benchmarks, but it often does not effectively utilize temporal information, and may struggle with discontinuous motion fields at organ boundaries. With this in mind, we develop and validate p4Dreg, an improved parameterized 4D registration algorithm tailored for 4DCT imaging.
Methods:
Five clinical abdominal 4DCTs were analyzed. Each scan was registered in a group-wise fashion, where all phases were registered toward the average anatomical position. A nnUNet segmentation model was applied to segment the body cavity to constrain the spatial smoothness calculation, minimizing deformation vector field (DVF) discontinuities at the cavity border. In addition, a new circular continuity regularization loss term was added such that the DVFs were continuously smooth from the first to the last phase of the breathing cycle. This helps discourage errors that may arise from image artifacts of certain phases.
Results:
The dice scores computed for 11 major abdominal organs were 0.97±0.02, while the average surface to the ground truth surface distance was 0.1±0.7 mm. The algorithm showed consistency in projections of vessel bifurcation landmarks, with an error standard deviation of 0.7mm, and no error greater than 4mm. These were excellent results compared to the abdominal DIR accuracy in the order of 9 to 38 mm reported in the literature. In addition, our results showed improved robustness compared to the high-performing algorithm pTVreg, which had a landmark error standard deviation of 1.2mm and a maximum error of 7mm on our testing data.
Conclusion:
Our algorithm demonstrated high accuracy and consistency in abdominal 4DCT data. The smoothness and robustness of the algorithm on the testing cases are critical for clinical implementation and avoiding costly errors.