Task-oriented compression of medical images.
Magnetic Resonancae Tomography produces large quantities of three-dimensional medical image data. Data compression techniques can be used to improve the efficiency with which these images can be stored and transmitted, but in order to achieve significant compression gains, lossy compression techniques (which introduce distortion into the images) must be used. Conventional metrics of distortion do not measure the effect of this "loss" on tasks applied to the images. This thesis uses a new task-oriented quality metric which measures the similarity between a radiologist's manual segmentation of brain lesions in raw (not compressed) magnetic resonance images and automated segmentations performed on raw and compressed images. To compress the images, a general wavelet-based lossy image compression technique, embedded zerotree coding, is used. A new compression system is designed and implemented which enhances the performance of the zerotree coder by using information about the location of important anatomical regions in the images, which are coded at different rates. Application of the new system to magnetic resonance images is shown to produce compression results superior to the conventional methods, with respect to the segmentation similarity metric.