PET image clustering

Image clustering is a method where image voxels are grouped by the similarity of their time courses (Ashburner et al., 1996). The volumes-of-interest (VOIs) are created automatically and objectively, instead of manual VOI drawing, and are based on functional, not anatomical, data. Clustering process may include spatial constraints, for example when used to delineate tissue lesions. Enhanced random walk algorithm for VOI delineation (Stefano et al., 2017) includes clustering step.

Clustering could be used to extract image-derived input function. Spatial information is not used in extraction of reference region curve, because in many brain diseases the pathology is widely spread throughout the entire brain, and an anatomically defined reference region cannot be found (Gunn et al., 1998). Tissue TACs with different kinetics are needed for model-based input function, and clustering could be used for delineation of ROIs for this purpose (Zheng et al., 2011). In diagnostic FDG brain studies, clustering can be used to extract the reference for ratio calculation (Yakushev et al., 2009; Borghammer et al., 2009; Dukart et al., 2013).

Clustering can also be applied in calculation of parametric images to reduce the computation times and to improve signal-to-noise ratio (Kimura et al., 1999; Bentourkia, 2001), or to enhance visualization of tumour tissue be removing surrounding healthy tissues with high signal but different kinetics (Gray et al., 2010). Compartmental model fitting and clustering can be combined to further enhance the quality of parametric image in case of very noisy data (Mohy-ud-Din et al., 2015).

The similarity of two time-activity curves should be not be based on sum-of-squares alone. Other straight-forward approaches are the Akaike Information Criterion (Kletting et al., 2009), runs test, and maximum run length (MRL) (Herholz et al., 1989).

The k-means method (Hartigan and Wong, 1979) remains as the most widely used clustering method because of its simplicity, but its performance is dependent on a good initial cluster. Its speed and accuracy can be improved with seeding techniques, such as k-means++. The k-means method combined with PCA seems to be the most promising candidate for organ-level clustering from dynamic total-body PET images (Jaakkola et al., 2023).

Affinity propagation (AP) method is more reliable than k-means and does not need initial clusters (Frey and Dueck, 2007). In AP, image voxels in each cluster are associated with an prototypical voxel (exemplar). Initially, all voxels are assumed to be exemplars. In each iteration, all voxels pass two values (messages) to all the other voxels: how responsible the voxel is to be the exemplar of the other voxel (responsibility), and how available the voxel is for serving as an exemplar (availability). Iterations are stopped when the messages stop changing (Frey and Dueck, 2007; Foster et al., 2014). AP-based segmentation can used for denoising (Xu et al., 2014), and be combined with PVC (Xu et al., 2018).

See also:


Abualhaj B, Weng G, Ong M, Attarwala AA, Molina F, Büsing K, Glatting G. Comparison of five cluster validity indices performance in brain [18F]FET-PET image segmentation using k-means. Med Phys. 2017; 44(1): 209-220. doi: 10.1002/mp.12025.

Acton PD, Pilowsky LS, Kung HF, Ell PJ. Automatic segmentation of dynamic neuroreceptor single-photon emission tomography images using fuzzy clustering. Eur J Nucl Med. 1999; 26(6): 581-590. doi: 10.1007/s002590050425

Ashburner J, Haslam J, Taylor C, Cunningham VJ, Jones T. A cluster analysis approach for the characterization of dynamic PET data. In: Quantification of Brain Function Using PET, Academic Press, 1996, pp 301-306.

Bankman IH (ed.): Handbook of Medical Image Processing and Analysis, 2nd ed., Academic Press, 2009. ISBN 978-0-12-373904-9. doi: 10.1016/B978-0-12-373904-9.X0001-4.

Bentourkia M. A flexible image segmentation prior to parametric estimation. Comput Med Imaging Graph. 2001; 25: 501-506. doi: 10.1016/S0895-6111(01)00016-7.

Boudraa A, Champier J, Cinotti L, Bordet J-C, Lavenne F, Mallet J-J. Delineation and quantitation of brain lesions by fuzzy clustering in positron emission tomography. Comput Med Imaging Graph. 1996; 20(1): 31-41. doi: 10.1016/0895-6111(96)00025-0.

Foster B, Bagci U, Mansoor A, Xu Z, Mollura DJ. A review on segmentation of positron emission tomography images. Comput Biol Med. 2014; 50(1): 76-96. doi: 10.1016/j.compbiomed.2014.04.014.

Gunn RN, Lammertsma AA, Cunningham VJ. Parametric imaging of ligand-receptor interactions using a reference tissue model and cluster analysis. In: Quantitative Functional Brain Imaging with Positron Emission Tomography. Academic Press, 1998, pp 401-406.

Herholz K, Heiss WD, Pietrzyk U, Wienhard K. Pixel-by-pixel fits of blood volume, transport, and metabolic processes: principle, normal values, and brain tumor studies with dynamic FDG-PET. In: Beckers C, Goffinet A, Bol A (eds.): Positron Emission Tomography in Clinical Research and Clinical Diagnosis: Tracer Modelling and Radioreceptors, pp 148-161. Kluwer, 1989. ISBN: 0-7923-0254-0.

Kimura Y, Hsu H, Toyama H, Senda M, Alpert NM. Improved signal-to-noise ratio in parametric images by cluster analysis. NeuroImage 1999; 9(5): 554-561. doi: 10.1006/nimg.1999.0430.

Kimura Y. Formation of parametric images with statistical clustering. Int Congr Ser. 2004; 1265: 25-30. doi: 10.1016/j.ics.2004.04.044.

Kletting P, Kull T, Reske SN, Glatting G: Comparing time activity curves using the Akaike information criterion. Phys Med Biol. 54: N501-N507, 2009. doi: 10.1088/0031-9155/54/21/N01

Liptrot M, Adams KH, Martiny L, Pinborg LH, Lonsdale MN, Olsen NV, Holm S, Svarer C, Knudsen GM. Cluster analysis in kinetic modelling of the brain: a noninvasive alternative to arterial sampling. NeuroImage 2004; 21: 483-493. doi: 10.1016/j.neuroimage.2003.09.058.

Mateos-Pérez JM, García-Villalba C, Pascau J, Desco M, Vaquero JJ. jClustering, an open framework for the development of 4D clustering algorithms. PLoS One 2013; 8(8): e70797. doi: 10.1371/journal.pone.0070797.

Mohy-ud-Din H, Lodge MA, Rahmim A. Quantitative myocardial perfusion PET parametric imaging at the voxel-level. Phys Med Biol. 2015; 60: 6013-6037. doi: 10.1088/0031-9155/60/15/6013.

Saad A, Smith B, Hamarneh G, Möller T. Simultaneous segmentation, kinetic parameter estimation, and uncertainty visualization of dynamic PET images. In: Ayache N, Ourselin S, Maeder A (eds): MICAAI 2007, Part II, LNCS 4792, pp 726-733, 2007.

Terry JL, Crampton A, Talbot CJ. Associating families of curves using feature extraction and cluster analysis. In: Algorithms for Approximation. Proceedings of the 5th International Conference, Chester, July 2005. Iske A, Levesley J (Eds.), Springer, 2007, pp 71-80.

Wong K-P, Feng D, Meikle SR, Fulham MJ. Segmentation of dynamic PET images using cluster analysis. IEEE Trans Nucl Sci. 2002; 49(1): 200-207. doi: 10.1109/TNS.2002.998752.

Tags: , ,

Updated at: 2024-01-15
Created at: 2015-02-25
Written by: Vesa Oikonen