CT Data Storage Reduction by Means of Compressing Projection Data Instead of Images: Feasibility Study 1

Transcription

1 Kyongtae T. Bae, MD, PhD Bruce R. Whiting, PhD Index terms: Computed tomography (CT), image display and recording Computed tomography (CT), image processing Computed tomography (CT), technology Data compression Radiology 2001; 219: From the Mallinckrodt Institute of Radiology, Washington University School of Medicine, 510 S Kingshighway Blvd, St Louis, MO From the 2000 RSNA scientific assembly. Received July 21, 2000; revision requested September 7; revision received October 3; accepted October 18. Address correspondence to K.T.B. ( baet@mir.wustl.edu). RSNA, 2001 Author contributions: Guarantors of integrity of entire study, K.T.B., B.R.W.; study concepts, K.T.B.; study design, K.T.B., B.R.W.; literature research, B.R.W.; clinical studies, K.T.B.; experimental studies, B.R.W.; data acquisition, K.T.B.; data analysis/interpretation, K.T.B., B.R.W.; manuscript preparation, definition of intellectual content, editing, revision/review, and final version approval, K.T.B., B.R.W. CT Data Storage Reduction by Means of Compressing Projection Data Instead of Images: Feasibility Study 1 The authors developed and evaluated a technique of compressing raw projection data at computed tomography (CT). Raw projection data acquired at CT were compressed and decompressed and then used for image reconstruction. For comparison, original images were compressed by comparable ratios. Projection data files were more compressible than image files. Projection data compression is a promising, efficient method to reduce data file size and thus to facilitate retrospective image reconstruction. Recent advances in computed tomographic (CT) technology, such as spiral scanning and multidetector-row technology (1,2), enable faster scanning of more patients and/or a larger volume of images for each patient, resulting in a continuous increase in digital data volume. Furthermore, advances in digital technology and algorithm development, such as picture archiving and communication systems, or PACS, allow interactive displays of larger amounts of image data. A serious challenge to implementing a picture archiving and communication system is the amount of data that must be accommodated. Image compression is a common approach to reduce the image data file size by means of transformations and encoding schemes to represent image information with fewer bits (3). Various schemes have been proposed and developed (4). One class of algorithms is called lossless compression, in which the information of the original image is exactly preserved. Lossless compression ratios are typically limited to 1:2 1:3. Techniques that allow loss of information during compression of the original image are known as lossy compression and can reach much higher compression ratios, with more compression generally resulting in higher distortion of the final image. Studies have shown that CT images with compression ratios in the range of 8:1 have observable distortions (4), which may affect the confidence of the interpretation by radiologists. One unique feature of CT compared with other imaging modalities is that the CT scanning process involves collecting raw projection data before reconstructing images. This collection of raw projection data is also termed a sinogram, because the data appear as a set of sinusoidal lines and bands (5). Sinusoidal structures represent x-ray beam attenuation signals that are generated from a rotating x-ray tube, transmitted through objects, and received by a ring or arc of detectors. These sinusoidal structures are highly correlated and may allow compression schemes to efficiently reduce the amount of stored data. The sinogram image is conventionally displayed in a coordinate system determined by the positions of the CT x-ray tube and the detector for the x and y coordinates, respectively. A sinogram contains information that is not available from its reconstructed image. For instance, a full CT scanning field of view is recorded in a sinogram, whereas a small subsection or region of interest of the full field of view is usually reconstructed in an image. At spiral CT, sinograms are acquired in a continuous fashion. With sinograms available, any transverse image can be specified for reconstruction at any longitudinal interval of scanning, which is particularly useful when overlapping images with a small reconstruction increment are needed to improve resolution (6). Despite their added value over images, sinogram data are not routinely stored because a sinogram is typically three to 850

2 Figure 1. Abdominal CT image, its original and compressed sinograms, and difference between these sinograms. A, Transverse abdominal CT image obtained with a standard, nonspiral technique with 8-mm collimation. B, Original uncompressed sinogram that consists of 1,252 tube positions in the x coordinate and 768 detectors in the y coordinate. C, Sinogram compressed at 12:1 compression ratio. D, Difference image between the original and compressed sinograms. It is difficult to distinguish the original and compressed sinograms by means of visual inspection without obtaining a difference image. Figure 2. Compressed abdominal CT images and their differences from the original uncompressed image. A, Image reconstructed from the 12:1 compressed sinogram in Figure 1, C. B, Image with 12:1 direct compression. C, Difference image between A and the original image in Figure 1, A. D, Difference image between B and the original image in Figure 1, A. The compressed images are not easily discernible from the original image; however, B shows a slight increase in low-level white noise. Differences are more pronounced on the direct-compression image than on the sinogram-compression image. The root-mean-square error, which quantifies the magnitude of noise, was and for images C and D, respectively. than image data because of their highly correlated structure. The objective of our study was to develop a technique of compressing sinogram data and to evaluate its performance in clinical CT cases. four times larger in data file size than an individual CT image. We postulate that sinogram data are more compressible Materials and Methods Clinical diagnostic images and their corresponding raw projection data, or sinograms, of the chest, abdomen, and head were acquired (Somatom Plus-4; Siemens Medical Systems, Iselin, NJ) in three adult patients. The chest projection scan was obtained in male patient aged 63 years with a spiculated lung nodule; the scan was obtained with a thin-section spiral CT technique (collimation, 2 mm; Volume 219 Number 3 CT Data Storage Reduction 851

3 pitch, 1). The abdominal scan was acquired in a female patient aged 45 years with a questionable liver mass by using a standard, nonspiral technique with 8-mm collimation. The head scan was acquired in a male patient aged 26 years with severe headache by using the same nonspiral technique. These abdominal and head scans were interpreted as normal. It was our intention to test the compression technique for both spiral and nonspiral scans. In our data set, the choice of spiral or nonspiral CT scans for a body part was arbitrary and based on the availability of scans at the time of data collection. The collected CT images and their sinograms were electronically transferred to a workstation (Sun Ultra 10; Sun Microsystems, Palo Alto, Calif). Sinograms consisted of either 768 or 1,536 detectors with 1,252 samples per gantry revolution. The data were extracted and processed into units representing the logarithm of the ratio of the unattenuated x-ray beam to the transmitted x-ray beam. Data compression was performed with an industry standard compression scheme (JPEG-LS implemented with LOCO; Hewlett-Packard, Palo Alto, Calif) (7). A low-complexity pattern generation and run-length encoding scheme for data representation are used in the compression scheme. The program executed encoding or decoding at a speed of approximately 1 Mbyte/sec. The amount of compression in the compression scheme was controlled by specifying the maximum allowable error between an original pixel value and the reconstructed value (eg, error 2 meant that all pixels in the final image are within two counts of their original value). Compression levels were specified by the maximum allowable code value error in a pixel, which produced multiple incremental compression ratios of 12:1, 20:1, and 23:1. The choice of these ratios was determined arbitrarily. By specifying the allowable error to be some multiple of the intrinsic noise in the CT acquisition process, the error introduced can be on the order of the underlying physical noise and, therefore, should be minimally objectionable (8). Sinograms were compressed and decompressed and then transferred back to the CT scanner console for image reconstruction. Also, images reconstructed from the original sinogram were compressed by compression ratios comparable with the sinogram compression and then decompressed. The compression ratio for a given specified error was found to depend on the details of the scanned object Figure 3. Head CT image, its original and compressed sinograms, and difference image between these sinograms. A, Transverse head CT image obtained with a standard, nonspiral technique with 8-mm collimation. B, Original uncompressed sinogram. C, Sinogram compressed at 12:1 compression ratio. D, Difference image between the original and compressed sinograms. The difference is most prominent at the interface between the head and surrounding air. Notice that the relative area (a band of bright zone flanked by arrows in B) occupied by the head cross section in the sinogram is smaller than that occupied by the abdominal cross section in Figure 1. being compressed; therefore, program parameters were varied to achieve comparable compression ratios for image comparison. Images from compressed sinograms were compared with those from direct image compression. For each of the three scans, image sets (consisting of an original image, images compressed at three levels, and images reconstructed with three levels of sinogram compression) were viewed on the scanner console workstation. The quality of these images was directly compared (K.T.B.) by using a side-by-side display. No systematic or blinded analysis to grade the image quality was attempted. Examined qualitative properties included low-level contrast in homogeneous portions of the image, edge distortion, and artifact visibility. In addition, a difference image between the original and compressed images was generated for each image data set and studied for correlated structure or artifacts. The root-mean-square error of the difference, which is an engineering metric commonly used for image degradation, was computed for a quantitative measure. Results A nonspiral abdominal CT image, the original and 12:1 compressed sinograms, and the difference images between these sinograms are shown in Figure 1. The x coordinate of the sinogram represents the loci of the x-ray tube, while the y coordinate represents the loci of the detectors. Each horizontal line of the sinogram represents transmitted x-ray beam signals recorded at each detector during one revolution of the CT gantry. The bright, white-cloud area of the sinogram corresponds to the abdomen in cross section, while the peripheral dark area corresponds to the surrounding air. Al- 852 Radiology June 2001 Bae and Whiting

4 Figure 4. Compressed head CT images and their differences from the original uncompressed image. A, Image reconstructed from 12:1 compressed sinogram in Figure 3, C. B, Image with 12:1 direct compression. C, Difference image between A and the original image in Figure 3, A. D, Difference image between B and the original image in Figure 3, A. The difference is more pronounced on the direct-compression image than on the sinogram-compression image. The root-mean-square error was 1.46 and 2.35 for images C and D, respectively. Peripheral halo areas (arrows) of distinct textural pattern in the difference images represent the air space surrounding the head region and are likely caused by reconstruction kernel clipping. Figure 5. Head CT images. A, Image reconstructed from 23:1 compressed sinogram. B, Image with 20:1 direct compression. Both transverse images show a substantial degradation of image quality compared with the original image in Figure 3, A. Marked low-level white noise is present on the sinogram-compression image, resulting in a substantial loss of the gray and white matter differentiation. However, the direct-compression image shows more severe image degradation with the presence of high-level structural noise and streaky artifacts. though it was difficult to distinguish the compressed from the original sinograms by means of visual inspection, the difference was most prominent at the boundaries of the abdomen, the CT table, and air because of the homogeneous texture of unattenuated x rays in air. An image reconstructed from the 12:1 compressed sinogram in Figure 1 and an image with 12:1 direct compression are shown in Figure 2, A and B, respectively. Difference images were generated between the original and the sinogram-compression images and between the original and the direct-compression images; these are shown in Figure 2, C and D, respectively. These compressed abdominal images were not easily discernible from the original image in Figure 1. Compared with the sinogram-compression image, however, the direct-compression image demonstrated a slight increase in low-level white noise, particularly in homogeneous areas of the subcutaneous fat and the liver. This increase in low-level noise was clearly demonstrated qualitatively in the difference images and quantitatively from the computation of the root-mean-square error of the difference images. The root-meansquare error that quantified the magnitude of noise was and in the difference images from the sinogram compression and the direct compression, respectively. No high-level structural noise or artifacts were introduced from the compression. A nonspiral head CT image, its original and 12:1 compressed sinograms, and the difference image between these sinograms are shown in Figure 3. The crosssectional area of the head was smaller than that of the abdomen, resulting in a larger area of air space surrounding the head on the CT scan. This large air space was clearly delineated on the sinograms. An image reconstructed from the compressed sinogram in Figure 3 and an image with 12:1 direct compression are shown in Figure 4, A and B, respectively. Although these images were difficult to discern from the original, the direct-compression image showed more increase in low-level white noise than did the sinogram-compression image. This increase in noise was confirmed on the difference images, as shown in Figure 4, C and D, and was quantified in the root-meansquare error measurements: 2.35 for the direct-compression image difference and 1.46 for the sinogram-compression image difference. The head image reconstructed from a 23:1 sinogram compression is shown in Figure 5, A. For comparison, a directcompression image was generated from Volume 219 Number 3 CT Data Storage Reduction 853

5 the original image with a 20:1 compression ratio (Fig 5, B). Both images were profoundly degraded from the original image. Marked increase in low-level noise was observed on the sinogram-compression image, resulting in a substantial loss of the gray and white matter differentiation. However, the direct-compression head image showed more severe image degradation and the presence of highlevel structural noise and streaky artifacts. An image from a spiral lung CT scan is shown with its spiral sinogram in Figure 6. At the same section level, an image was reconstructed from 12:1 sinogram compression in a lung window setting (Fig 7, A) and a soft-tissue window setting (Fig 7, C). For comparison, a 12:1 direct-compression image was generated with a lung window setting (Fig 7, B) and a soft-tissue window setting (Fig 7, D). The images displayed with the soft-tissue window setting showed slightly more increase in low-level noise on the direct-compression image than on the sinogram-compression image, particularly in the heart and subcutaneous fat regions. Figure 6. Spiral lung CT image and its sinogram. A, Transverse thin-section lung CT image obtained with a spiral technique, with 2-mm collimation and pitch of 1. B, Original uncompressed sinogram with multiple interweaving sinusoidal paths that are characteristics of the spiral CT sinogram. Discussion One of the important practical features of spiral CT is that retrospective reconstruction can be performed. When spiral raw projection data are stored instead of reconstructed images, these data can be recalled later to generate images with different reconstruction increments, depending on clinical applications. Similarly, in multidetector-row spiral CT, images with different section thicknesses can be generated from the same set of raw projection data when needed. Only the image data are routinely stored, however, because the size of the projection data file is usually as large or larger than the total reconstructed image sections. The original projection data that are used to reconstruct image sections are frequently kept for only a day or so, in case further views are needed for diagnosis. If an efficient data compression technique exists, having the original projection data available would provide the additional benefit of allowing additional viewing and processing at a later date. As an example to demonstrate the savings in data file size with compression, the original projection data of a standard spiral abdominal pelvic CT study are approximately 115 Mbyte in size, while the compressed file size is less than 10 Mbyte at 12:1 compression. A set of 60 recon854 䡠 Radiology 䡠 June 2001 Figure 7. Compressed spiral lung CT images displayed with lung (A, B) and soft-tissue (C, D) window settings. A, Image reconstructed from 12:1 compressed sinogram in Figure 6, B. B, Image with 12:1 direct compression. C, Image A with a soft-tissue window setting. D, Image B with a soft-tissue window setting. The images displayed with the soft-tissue window setting show slightly more increase in low-level noise on the direct-compression image than on the sinogramcompression images, particularly in the heart and subcutaneous fat regions. structed images in such a study would be about 32 Mbyte. The sinogram can correspond to a stationary scan of a single plane (leading to a step and repeat acquisition process for the full volume) or the continuous movement of the object during the spiral CT acquisition. Because sinograms are formed by a projection (integrating) process, their constituent signals have a high degree of correlation and, hence, are amenable to various compression techniques. Reconstruction processes, such as filtered backprojection or simultaneous algebraic reconstruction, produce planar image sections from the sinogram data. While the sinogram contains information about all objects in its field of view (on the order of 50 cm), typically only a small area is of interest, so after a preview, a subset region of interest is reconstructed and ultimately saved. Thus, schemes that compress images may have the advantage that they address smaller, more correlated data sets but at the expense of loss of information about the whole view of the patient. The scope of our study was limited to Bae and Whiting

6 evaluating the technical feasibility of compressing projection data, rather than comparing different compression methods. We used an industry standard compression scheme, thereby providing ease of use and benefits of interoperability. By compressing in the sinogram domain, study information can be compressed substantially and rapidly, reducing system requirements while maintaining clinical utility. It may be useful to test other available compression algorithms, such as a wavelet-based compression method, or design a new compression algorithm that is more specific and perhaps more efficient in compressing sinograms. For example, compression techniques, such as differential pulse-coded modulation, efficiently encode data needed for eventual image reconstruction because reconstruction algorithms typically use filtered backprojection. In this study, no systematic analysis was performed to investigate the effect of compression ratios on the image quality. Compression ratios depended on the type of object being scanned. Compression ratios between 10:1 and 15:1 were found to be clinically acceptable, whereas higher compression ratios compromised diagnostic usefulness. At comparable compression ratios, images from the compressed sinograms were similar in quality to images that had been directly compressed, even though the compressed images covered only one-half to one-fourth of the total area of the sinogram. Artifacts introduced by compressing sinograms tended to be low-level white noise and, therefore, degraded high-contrast regions less. Although our current implementation was limited to the data from a single CT vendor, our method can be easily applicable to sinograms obtained with other CT scanners, provided that header and scanner control information of the sinogram can be separated from projection measurement and that sinogram data are accessible in normalized form. In conclusion, we developed a technique of compressing CT projection data and evaluated its performance by comparing it with a direct image compression technique in several clinical cases. CT projection data appeared more compressible than image data; thus, a greater reduction in overall data file size could be achieved by compressing the projection data. Compression of projection data is a promising, efficient method to reduce data file size and to facilitate retrospective image reconstruction. References 1. Berland LL, Smith JK. Multidetector-array CT: once again, technology creates new opportunities. Radiology 1998; 209: Hu H, He HD, Foley WD, Fox SH. Four multidetector-row helical CT: image quality and volume coverage speed. Radiology 2000; 215: Rabbani M, Jones PW. Digital image compression techniques. In: Tutorial texts TT07. Bellingham, Wash: SPIE Press, Foos DH, Muka E, Slone RM, et al. JPEG 2000 compression of medical imagery. Proc SPIE Med Imaging 2000; 3980: Kak AC, Slaney M. Principles of computed tomography. New York, NY: IEEE Press, Kalendar WA, Polacin A, Suss C. A comparison of conventional and spiral CT: an experimental study on the detection of spherical lesions. J Comput Assist Tomogr 1994; 18: Weinberger M, Seroussi G, Sapire G. The LOCO-1 lossless image compression algorithm: principles and standardization into JPEG-LS. Palo Alto, Calif: Hewlett-Packard Laboratories, Young S, Whiting BR, Foos DH. Statistically lossless image compression for CR and DR. Proc SPIE 1999; 3658: Volume 219 Number 3 CT Data Storage Reduction 855