Abstract
1427
Objectives Traditionally, iterative fitting (IF) based compartment models may encounter problems when handling noisy data like the overfitted compartment parameters and lack of reproducibility. Therefore, there is need for a robust and reproducible solution, which can be applied to any kind of PET data by persons even without a long term experience.
Methods Here we proposed an innovative machine learning (ML) based kinetic modeling, which can predict the values of quantitative PET parameters using learned regression models from accumulated training studies. ML method can fully utilize the history knowledge to build a moderate kinetic model that can directly process noisy data. Also due to the database, our method can automatically adjust the regression models without much user interference using our multi-threads grid parameter searching technique. Furthermore, we use the candidate concept to combine the advantages of ML and IF modeling methods, which can find a balance between fitting to history data and unseen target curve.
Results We built a 2-tissue compartments modeling database for reference that consists of 502 VOI studies from 137 patients by kinetic modeling experts. We compared the IF and ML methods from the statistical point of view and biological aspects. [TABEL] The result showed that two methods have a similar statistical performance. However, ML only produced 3 cases of overfitting in contrast to 122 (24%) cases of IF method.
Conclusions ML method achieved a comparable accuracy as with the IF method. But it is more robust and stable without much invalid parameters when handling noisy data.
- © 2009 by Society of Nuclear Medicine