Segmentation of Spinal Disc Tissue in MR using Kalman Filters and Active Contour Models

We present a new technique for extracting disk tissue from sagittal MR spine images. This technique uses linear Kalman filters to estimate initial conditions for a two-dimensional active contour model. We extend the contour model by computing an energy penalty proportional to the mismatch between a generic model and the patient specific model (using a chi-squared error measure). We compared the output of our system to human-guided manual segmentation. We performed 30 experiments (using a T1 weighted 3DFGRE pulse sequence), varying the number and location of disk boundary seedpoints as input to the Kalman filter. The output of the filter initialized the active contour models for the disks to be segmented. For each experiment, we measured the average RMS error between the computed and human-detected boundaries. In all cases, the errors were less than 0.25 mm and the entire disc body could be extracted in under 10 minutes. Our results demonstrate that Kalman filters can be used to guide an active contour model to extract disk boundaries in a quick and robust manner.