Ilona Kowalik-Urbania | Applied Math, University of Waterloo
The quest for 'diagnostically lossless' medical image compression using objective image quality measures
Given the explosive growth of digital image data being generated, medical communities worldwide have recognized the need for increasingly efficient methods of storage, display and transmission of medical images. For this reason lossy image compression is inevitable. Furthermore, it is absolutely essential to be able to determine the degree to which a medical image can be compressed before its “diagnostic quality” is compromised. This work aims to achieve “diagnostically lossless compression”, i.e., compression with no loss in visual quality nor diagnostic accuracy.
Recent research by Koff et al. has shown that at higher compression levels lossy JPEG is more effective than JPEG2000 in some cases of brain and abdominal CT images. We have investigated the effects of the sharp skull edges in CT neuro images on JPEG and JPEG 2000 lossy compression. We provide an explanation why JPEG performs better than JPEG2000 for certain types of CT images.
Another aspect of this study is primarily concerned with improved methods of assessing the diagnostic quality of compressed medical images. In this study, we have compared the performances of structural similarity (SSIM) index, mean squared error (MSE), compression ratio and JPEG quality factor, based on the data collected in a subjective experiment involving radiologists. An receiver operating characteristic (ROC) curve and Kolmogorov-Smirnov analyses indicate that compression ratio is not always a good indicator of visual quality. Moreover, SSIM demonstrates the best performance. We have also shown that a weighted Youden index can provide SSIM and MSE thresholds for acceptable compression.
We have also proposed two approaches of modifying L2-based approximations so that they conform to Weber’s model of perception. We show that the imposition of a condition of perceptual invariance in greyscale space according to Weber’s model leads to the unique (unnormalized) measure with density function ρ(t) = 1/t. This result implies that the logarithmic L1 distance is the most natural “Weberized” image metric. We provide numerical implementations of the intensity-weighted approximation methods for natural and medical images.