• Jiang, Zhenzhen
  • 01.01.2011
  • 21.11.2011
In the last decades there has been a remarkable increase in the amount of digital data acquired by radiological high-resolution imaging modalities like computed tomography (CT). Currently this data is usually stored in uncompressed raw files, with lossless compression methods, or with widely accepted lossy compression standards like JPEG 2000. In the latter case, there have been quality assessments among radiologists, physicians and clinical imaging professionals, suggesting fixed data-rates or signal-to-noise ratios (SNRs) for different examinations or acquisitions. However, the results of surveys using such simple approaches can not easily be generalized to arbitrary image distortions occurring in more recent or data-adapted com-pression algorithms. There exist more elaborate metrics for automatic estimation of visual quality (e. g. Structural Similarity, SSIM), the visibility of distortions (Visual Discrimination Models, VDMs) and detectability of features like tumors (Model Observers).

The task of Ms. Zhenzhen Jiang is to evaluate the applicability of these metrics to compressed CT datasets. Of special interest with regard to different application areas are SSIM, Sarnoff model as an example of a VDM, and the well established (channelized) Hotelling Observer. These shall be compared to straight-forward error distance metrics ( ).To do so, in a first step the mentioned metrics shall be implemented in well-documented Matlab functions. With these implementations the features of the metrics, as well as their responses to different coding artifacts (e. g. arising from Gaussian noise, JPEG 2000, or motion compensation) shall be analyzed analytically and experimentally. Optionally, JPEG 2000 data-rates, suggested in human observer experiments for various examination procedures, can be translated to these metrics. In the course of the project thesis there is a collaboration with Mr. Michel Bätz, who implements voxel-based motion compensation for CT datasets, as well as with Mr. David Müller, who compares the performance of different codecs for CT datasets.