Data-driven motion compensation in HR-pQCT

High-resolution peripheral quantitative computed tomography (HR-pQCT) enables the assessment of bone mineral density and three-dimensional microstructure of peripheral limbs. Due to its long scanning time, it is especially susceptible to patient induced motion artifacts. The goal of this study is to compensate for such artifacts without requiring excessive computation time, while still avoiding the hallucination of erroneous structures. Approach. This work proposes a machine learning-based approach for the rapid estimation of rigid motion parameters directly from raw projection data. The model is trained on patient scans with motion simulated as a single, instantaneous jump. We demonstrate the performance and generalization of the method on simulated and experimentally acquired data, benchmarking our results against two recent baselines. Main results. The proposed methods greatly reduced the amount of motion artifacts in both simulated and experimental data. On the simulated test data, the median structural similarity index measure (SSIM) for the reconstruction results could be improved from 0.62 to 0.88. Similarly, the SSIM for the experimental data improved from 0.58-0.71 to 0.74-0.78. At the tibia site that was used for training, motion artifacts were successfully reduced to such an extent that 92 % of previously inaccessible scans could be reconsidered for clinical evaluation. For the unseen radius site, a clinical recovery rate of 56 % could be achieved. Predicting the motion parameters for a measurement required less than 200 ms on consumer hardware. Significance. The proposed approach has the potential to enable the reuse of scans that were previously discarded for clinical evaluation while preventing generative hallucinations by design. Consequently, eliminating the need for repeated scans would reduce patient radiation exposure and clinical workflow overhead.