Lossy Compression of Ultrasound Images: a Space-Frequency Approach
Ultrasound is a useful and popular medical imaging modality. Due to the large volume of images generated, storage and communications issues motivate research into ultrasound image compression. Lossy techniques achieve higher compression than lossless methods, but must preserve diagnostically relevant information and maintain acceptable aesthetic quality. This thesis applies the technique of space-frequency segmentation using balanced wavelet packet trees to the lossy compression of medical ultrasound images. In order to represent an image, space-frequency segmentation searches a large family of possible space-frequency partitions, together with sets of possible quantizers, and finds the rate-distortion optimal combination.
Zerotree coding is a leading high performance lossy compression technique, but because it uses a fixed frequency decomposition topology designed for natural images and allocates bits based only on coefficient magnitude, it produces distinctive artifacts at higher compression rates when applied to ultrasound images.
Space-frequency segmentation has not been previously applied to medical ultrasound images, and unlike other published results, real entropy coders are implemented to code the space-frequency representation. Space-frequency segmentation improves both objective and subjective quality when compared against zerotree coding using Set Partitioning In Hierarchical Trees (SPIHT), and since the technique is not yet mature, there remain many possibilities to further enhance the performance.
The improvement in subjective quality using space-frequency segmentation is confirmed by a study in which ultrasound radiologists assess the quality of compressed images with respect to diagnostically relevant features. The optimum space-frequency partition highlights the characteristic speckle texture of ultrasound images as a particularly difficult feature to compress, and leads to a more effective decomposition topology for preserving speckle.