Contributions to the quantitative analysis of dynamic PET studies using clustering approaches

• Autores: José María Mateos Pérez
• Directores de la Tesis: Juan José Vaquero López (dir. tes.)
• Lectura: En la Universidad Carlos III de Madrid ( España ) en 2014
• Idioma: inglés
• Tribunal Calificador de la Tesis: Jorge Ripoll Lorenzo (presid.), María Jesús Ledesma Carbayo (secret.), Steven Bachrach (voc.)
• Materias:
  • Ciencias médicas
    • Medicina del trabajo
      • Medicina nuclear
• Enlaces
  • Tesis en acceso abierto en: e-Archivo
• Resumen

  • Dynamic positron emission tomography (PET) is a widespread medical imaging technique that allows the quantification of different physiological parameters within the body and yields more information that the one provided by a single, static image. Quantification of these studies involves obtaining the input function, that is, the amount of tracer present in arterial blood at any given point in time, and the tissue time-activity curve (TAC) for the tissue or organ under study. The subjacent biological processes are modelled as the tracer exchange rates between the arterial activity source and a compartmental model; this mathematical approach allows to quantify different biological aspects (metabolic rates, blood flow, specific receptor binding) in a non-invasive way. Typically, arterial and tissue TACs are extracted from the image data by drawing a ROI over the areas of interest, either over the PET image or over some anatomical imaging modality, such as CT, and in some cases acquire some blood samples to correct the input function for metabolites, partial volume effects or other different sources of distortion that may bias the final result. While this ROI delineation is done normally by an experienced operator, this process is very slow and, more importantly, subjective and non-replicable. Furthermore, ROI delineation over registered anatomical images may group together regions that look identical in the CT image but have different underlying kinetics. These reasons have motivated the development of automatic segmentation or TAC extraction algorithms, of which there are several examples in the medical imaging literature. Most of the proposed methods involve the use of unsupervised machine learning algorithms or the direct application of dimensionality reduction techniques, such as PCA or SVD. This thesis studies the feasibility of supervised algorithms to extract the activity curves of dynamic studies based solely on the knowledge acquired about the kinetics of similar ones. Our experiments on three swine studies showed that the segmentation was successful and the obtained TACs allowed the computation of the kinetic analysis and obtained smaller errors in the kinetic parameters obtained from the mathematical model than the manual segmentations. Said supervised algorithms are not common in the literature but we have shown that they can be a viable option for very specific subset of cases. One of the problems of the published automatic segmentation algorithms is the general lack of published source codes or even binary distributions. As has been studied in the literature, this presents a problem by itself, as it forces other researchers to re-implement said algorithms. This work presents the development of an open framework for dynamic imaging clustering that includes the most commonly used algorithms and that can be easily extended by third parties through the use of its public API. The code for said framework has been published with a free software license to allow it to be modified by external researchers and adapt it to their needs. It has been developed as an ImageJ plugin to take advantage to all the imaging analysis functionalities already presented in said platform. Using this framework, we also present an improvement of the classical leader-follower algorithm. This unsupervised algorithm groups image voxels with similar TACs according to a threshold set by the user and creates as many clusters as necessary to form homogeneous regions. Due to the nature of the partial volume distortions that need to be removed from the final TACs as much as possible, the proposed method implements a two-step leader-follower modification. In this case, the image voxels are clustered according to both a similarity metric and a distance metric; particularly, the cosine similarity and the Euclidean distance were chosen for our tests. This algorithm successfully segmented all of the evaluated 24 mice imaging studies, yielding quantitative parameters after the kinetic modelling that were not significantly different from those obtained via manual delineation and maintained the differences between the three tracers used in this experiment. --------------------------------------------------------