Low Dose CT Perfusion using K-Means Clustering

We aim at improving low dose CT perfusion functional parameters maps and CT images quality, preserving quantitative information. In a dynamic CT perfusion dataset, each voxel is measured T times, where T is the number of acquired time points. In this sense, we can think about a voxel as a point in a T-dimensional space, where the coordinates of the voxels would be the values of its time attenuation curve (TAC). Starting from this idea, a k-means algorithm was designed to group voxels in K classes. A modified guided time-intensity profile similarity (gTIPS) filter was implemented and applied only for those voxels belonging to the same class. The approach was tested on a digital brain perfusion phantom as well as on clinical brain and body perfusion datasets, and compared to the original TIPS implementation. The TIPS filter showed the highest CNR improvement, but lowest spatial resolution. gTIPS proved to have the best combination of spatial resolution and CNR improvement for CT images, while k-gTIPS was superior to both gTIPS and TIPS in terms of perfusion maps image quality. We demonstratek-means clustering analysis can be applied to denoise dynamic CT perfusion data and to improve functional maps. Beside the promising results, this approach has the major benefit of being independent from the perfusion model employed for functional parameters calculation. No similar approaches were found in literature.