Four-Dimensional Similarity Filter for Cerebral CT Perfusion Maps in Acute Stroke
Sjoerd A M Tunissen1, Ewoud J Smit1, Mikhail Mikerov1
1Department of Medical Imaging, Radboudumc, Geert Grooteplein Zuid 10, 6525 GA Nijmegen, the Netherlands.
None:
Background Brain CT perfusion is used for diagnosing ischemic stroke and treatment planning. Purpose To optimize the four-dimensional similarity filter (4DSF), a noise-reducing filter, for CT perfusion in patients with acute stroke. Materials and Methods To reduce the noise of dynamic perfusion images, the 4DSF averages voxels with the most similar time-intensity curves. The commercial implementation searches for similar voxels based on temporal mean values. The proposed stroke-specific 4DSF (s4DSF) searches for voxels with similar peak times, thus measuring changes in blood supply that are characteristic of acute stroke. For evaluation, nine stroke scenarios were simulated using a digital phantom. For each scenario, a 30-phase CT perfusion protocol was simulated. Perfusion analysis was performed using (a) Bayesian estimation with 4DSF (clinical standard) and s4DSF and (b) singular value decomposition (SVD) on original (clinical standard) CT scans and those with s4DSF to obtain five different perfusion maps. For all maps, the contrast (difference in mean value) and root mean square error of the stroke and contralateral regions were compared with and without s4DSF. Improvements in contrast correctness (difference with contrast of reference) were evaluated with a binomial test. To show clinical potential, CT perfusion scans of 12 patients with a suspected stroke were retrospectively retrieved. In these cases, s4DSF performance was visually assessed and compared with the clinical Bayesian and SVD methods. Results Compared with the clinical version, the s4DSF improved contrast correctness in 64% (29 of 45) of maps derived with Bayesian estimation and 73% (33 of 45) of maps derived with SVD (P = .036 and P = .001, respectively). Moreover, the root mean square error was lower for the maps with s4DSF in 70% (63 of 90) of Bayesian- and 99% (89 of 90) of SVD-derived perfusion regions. Analysis of patient data yielded similar results, with s4DSF showing more visible and clearly delineated stroke regions. Conclusion The s4DSF shows better performance than current clinical filters in phantom simulations and patient cases. Therefore, application of the s4DSF could improve diagnostic accuracy of cerebral perfusion maps. © RSNA, 2025 Supplemental material is available for this article. See also the editorial by Krainik in this issue.
