A clustering approach to 4D MRI retrospective sorting for the investigation of different surrogates

Abstract Purpose In retrospective 4-Dimensional Magnetic Resonance Imaging (4D MRI) sorting, respiratory surrogate selection affects the image quality of reconstructed volumes. We propose a method for retrospective 4D MRI sorting based on clustering, which allowed us to compare the performance of single or multiple internal surrogates vs. a conventional external signal. Methods A k-medoids clustering algorithm was exploited for sorting 2D MRI into 4D MRI, relying on (A) multiple or (B) single automatically tracked internal landmarks or (C) respiratory belt signal. 4D MRI reconstructions for seven liver cancer patients were compared to those of the state-of-the-art mutual information (MI) approach. Sorting artifacts were measured by the root mean square error (RMSE) between the diaphragm profile and a fitted second order curve. Diaphragm and tumor motions were evaluated. Results The median RMSEs ranged 0.97–1.66 mm, 1.24–1.89 mm, 1.43–2.27 mm, 1.74–3.72 mm for the MI, (A), (B) and (C) methods, respectively. Significant differences (Friedman, α = 5%) were found between (C) and all other methods, and between (B) and MI approaches. The discrepancies between (A) and MI approaches ranged 1.1–6.2 mm and 0.7–5.3 mm respectively in diaphragm and tumor motions. Methods (A) and (B) showed similar ranges of motion. Conclusion With multiple internal points, our method yielded the description of a higher range of motion and similar image quality with respect to the MI approach. The single point method led to more artifacts, suggesting the superior suitability of multiple internal surrogates for retrospective 4D MRI sorting. Considering internal rather than external information favored superior performance.