Compartmental Modeling in Positron Emission Tomography: A model selection approach

摘要

Compartmental Modeling is used in Positron Emission Tomography (PET) and it is an important tool for analysis and kinetic studies of living systems. Its application allows doctors and radiologists to provide diagnosis and treatment for several diseases (e.g. heart ischemia) by image processing, representing a non invasive way to quantify biochemical and physiological processes. Because of a large number of compartment models available, the task to choose the most suitable in a statistical sense may be difficult sometimes. The current work presents an assessment method for compartmental models for cardiology studies using Information Criterion approach for simulated experiments, being helpful to start a model development. The methodology consists of statistical assessment of one, two and three compartments models using features obtained from experimental data. It enables to analyze and make a decision about the most suitable number of compartments to be applied in a particular clinical exam or study. Synthetic curves were created to test estimation task and model choice was made using Akaike's Information Criterion. Fitting curve procedure employs Levenberg- Marquardt and Nelder-Mead optimizations techniques with sensitivities equations approach. Signal-noise ratio of tracer concentrations curves were estimated from experimental data (5 patients from Heart Institute of Medicine School of University of Sao Paulo, Brazil) and considered to be Gaussian in all simulated cases. Conclusion: Identification process was tested successfully for simulated data. For one and two compartments structures, 60 measures were enough to distinguish what model was employed to synthesize the respective data thanks to Akaike's Information Criterion. However, for three compartments simulated data, 200 points were necessary. This result shows that one needs more measures for complex models identification. Sampling carefully is a must when using three compartment models for 60 minutes scans-n, for instance. The next step consists of identification of real exams using commercial software as gold standard.