Comparison of pattern classification methods in segmentation of dynamic PET brain images
Heidi Koivistoinen, Jussi Tohka, Ulla Ruotsalainen · 2004
In this study, pattern classification methods for automatic segmentation of PET brain receptor density images are compared. Because of low contrast to noise ratio we utilize information about dynamics of tracer uptake present in PET studies. We compare three methods: expectation-maximization algorithm (EM), fuzzy C-means algorithm (FCM) and independent component analysis (ICA). Particularly, our interest was in segmentation of the striatum and cerebellum structures. The methods were applied to a Monte Carlo simulated phantom image and to five different human PET studies. We were able to extract striatum with the EM algorithm and ICA satisfactorily from all PET studies. With the FCM algorithm striatum could not be differentiated to its own class. The cerebellum was found only with ICA from the simulated image. ICA seemed to be less sensitive to noise of all the studied methods. The EM algorithm was most sensitive to patient movement in the human PET studies. The EM algorithm and ICA seemed promising for the segmentation task when taking low contrast to noise ratios of the PET images into account.