Abstract
1320
Objectives Evaluate the feasibility of ICA minimizing the mutual information (MI) for extraction of time activity curves (TACs) from dynamic PET images. Optimization is performed in polar coordinates to visualize the MI criteria and to examine the effects of the joint histogram bin size.
Methods In the algorithm, each pixel is represented as a vector. Principal component analysis and data whitening is first performed. The directions of ICA weight vectors is then optimized by non-orthogonal rotations in multidimensional space minimizing the total MI between each of these weight vectors. The effect of joint histogram bin size (128x128, etc) on the MI function was evaluated. Generation of the TAC and ICA images were performed by inversion of the final transformation matrices. Various types of dynamic data were tested: digital phantoms with spatially independent and dependent (mixed) data, Rb-82 cardiac, and 18F-FDG brain.
Results The MI as a function can have many local minima if calculated with a joint histogram of less than 256x256 bins. ICA extracted the shape and magnitude of the underlying TACs and component images of the unmixed phantom data. In spatially dependent data, the MI criteria necessitated additional constraints (positive vector elements). The method generated ICA images and TACs of arterial, venous and tissue phases the heart and brain images. The FDG brain data showed the carotid curves to be similar to the actual plasma sampled activity curves, when accounting for partial volume effect.
Conclusions Signal separation by ICA using direct minimization of the MI criteria is feasible and no a priori model density function is needed. Additional constraints are needed for images with large areas of spatially dependent signals. Polar coordinates provide a visual check on the solution and show a smoother MI function using 256 x 256 joint histogram bins. Further studies will be needed to determine the utility of this method for tracer kinetic modeling