T.Grenier, C. Muller
Purpose and Context
Image segmentation remains an arduous task due to the existence of various kinds of image noises. The objective is to find the accurate information that characterizes each object in the image.
Currently, most of the noninvasive medical imaging systems are able to provide several images of the same object:
- in PET: images resulting from various tracers (such as [18F]-NaF and FDG)
- in MRI: images of contrast weighted by r, T1, T2; cartographies of the parameters r, T1, T2.
- in ultrasonic imaging: envelope signal, density of scatterers, echogenicity, RF signal
- …
In this context we develop filtering and segmentation methods taking into account the whole information, in order to benefit from the complementarities of the parameters and achieve a more robust segmentation.
We specifically explore the Mean Shift procedure derived from the theory of nonparametric density estimation, using a multidimensional kernel. This approach is also based on a multidimensional feature space which is particularly appropriate for representing the whole set of observations, regardless of their types (scalar, vector, …).
Methods
We have proposed three methods:
- VBMS (Variable Bandwidth Mean Shift) : A fully adaptive Mean Shift procedure based on the impulse response of imaging system
- MPMS : A framework including a multiparametric Mean Shift filtering and an image segmentation method (based on region growing)
- Hybrid MPMS : A new iterative Mean Shift approach improving the MPMS filtering. Hybrid MPMS combines two Mean Shift procedures called nonblurring and blurring
Figure 1 :Hybrid Mean Shift flowchart
Results
We applied positively our methods on ultrasonic simulated images. Here are some results of MPMS and the hybrid MPMS methods on multiparametric data. These multiparametric data consist of three ultrasonic parameters: the envelope image, the estimation of scatterers density and a parameter of the Nakagami distribution.
We have assessed the improvements of Hybrid MPMS compared with the simple MPMS filtering
Figure 2 : Results of MPMS and Hybrid filtering on simulated US images
[GREN-05] T. Grenier. Apport de l'espace des caractéristiques et des paramètres d'échelle adaptatifs pour le filtrage et la segmentation d'image. Doctorat, INSA Lyon, December 2005. Jury: P. Refregier ( prés.), D. Comaniciu (rap.), J.M. Nicolas (rap.),M. Jourlin (rap.), C. Revol-Muller (co-dir.), G. Gimenez (dir.).