TREE-STRUCTURED VECTOR QUANTIZATION OF CT CHEST SCANS - IMAGE QUALITYAND DIAGNOSTIC-ACCURACY

The quality of lossy compressed images is often characterized by signa l-to-noise ratios, informal tests of subjective quality, or receiver o perating characteristic (ROC) curves that include subjective appraisal s of the value of an image for a particular application. We believe th at for medical applications, lossy compressed images should be judged by a more natural and fundamental aspect of relative image quality: th eir use in making accurate diagnoses. We apply a lossy compression alg orithm to medical images, and quantify the quality of the images by th e diagnostic performance of radiologists, as well as by traditional si gnal-to-noise ratios and subjective ratings. Our study is unlike previ ous studies of the effects of lossy compression in that we consider no n-binary detection tasks, simulate actual diagnostic practice instead of using paired tests or confidence rankings, use statistical methods that are more appropriate for non-binary clinical data than are the po pular ROC curves, and use low-complexity predictive tree-structured ve ctor quantization for compression rather than DCT-based transform code s combined with entropy coding. Our diagnostic tasks are the identific ation of nodules (tumors) in the lungs and lymphadenopathy in the medi astinum from computerized tomography (CT) chest scans. Radiologists re ad both uncompressed and lossy compressed versions of images. For the image modality, compression algorithm, and diagnostic tasks we conside r, the original 12 bit per pixel (bpp) CT image can be compressed to b etween 1 bpp and 2 bpp with no significant changes in diagnostic accur acy. The techniques presented in this paper for evaluating image quali ty do not depend on the specific compression algorithm and are useful new methods for evaluating the benefits of any lossy image processing technique.