Implementation of k-means clustering algorithm in 177Lu-dotate dosimetry calculations

Research output: Contribution to journalConference articleScientific


Aim: Single photon emission computed tomography (SPECT) enables quantitative in vivo measurements of radiopharmaceutical distributions, which are used for radionuclide therapy (RT) dosimetry calculations. Acquiring time activity curves (TAC) manually from a set of SPECT images appears operator-dependent
and time-consuming. Using semi- or fully-automated volume of interest (VOI) delineation methods the measurements can be performed more objectively and the reproducibility and the coherence of the dosimetry calculation process can be improved. In this study, we investigate the use of an unsupervised learning clustering algorithm built for TAC analysis and dosimetry calculations.

Materials and methods: A male patient received 7.4 GBq of 177Lu-dotatate
with kidney protective amino acids infusion. Four separate SPECT/CT acquisitions were performed at 1 h, 26 h, 3 d and 7 d post injection. Separate SPECT images were coregistered, filtered and masked with a simple threshold method to restrict
the analysis within the patient volume. A clustered composition of the images was created using a k-means clustering algorithm. The clustering was run using several different number of clusters (K) and the results were evaluated by visual assessment and calculating average mean-squared-error (MSE) across the
clusters. Average TACs were extracted from the cluster centroids.

Results: The average MSE decreases as a function of K, as expected. At K > 10, the MSE reaches a plateau and decreases slowly towards zero. Visually, homogeneous regions fragmented when using more than eight (K = 8) clusters. Interestingly, most of the hot spots in the patient’s liver and both kidneys were
separated from background even with three clusters. The next step is to calculate the absorbed doses corresponding to the different number of K –values and the different protocols used in the clinic.

Discussion and Conclusion: A simple K-means clustering algorithm provides an easy access to TAC of 177Lu-dotate SPECT studies. A moderate number of clusters is recommended to provide coarse TACs for different tissues. The clustered image (VOI-image) might also be useful for data comparison, when the patient goes through several separate treatment cycles. In order to plan the treatment and to assess the response advantageously, it is essential to perform the dosimetry calculations on a patient specific basis and it appears to us that the clustering algorithms may provide an effective tool to improve this process in the near future.
Original languageEnglish
JournalEuropean Journal of Nuclear Medicine and Molecular Imaging
Issue numberSuppl 2
Pages (from-to)S311
Number of pages1
Publication statusPublished - 2012
MoE publication typeB3 Article in conference proceedings
EventAnnual Congress of the EANM 2012 - , Italy
Duration: 1 Jan 1800 → …

Fields of Science

  • 114 Physical sciences

Cite this