Determining optimal medical image compression: psychometric and image distortion analysis
© Flint; licensee BioMed Central Ltd. 2012
Received: 29 March 2012
Accepted: 17 July 2012
Published: 31 July 2012
Storage issues and bandwidth over networks have led to a need to optimally compress medical imaging files while leaving clinical image quality uncompromised.
To determine the range of clinically acceptable medical image compression across multiple modalities (CT, MR, and XR), we performed psychometric analysis of image distortion thresholds using physician readers and also performed subtraction analysis of medical image distortion by varying degrees of compression.
When physician readers were asked to determine the threshold of compression beyond which images were clinically compromised, the mean image distortion threshold was a JPEG Q value of 23.1 ± 7.0. In Receiver-Operator Characteristics (ROC) plot analysis, compressed images could not be reliably distinguished from original images at any compression level between Q = 50 and Q = 95. Below this range, some readers were able to discriminate the compressed and original images, but high sensitivity and specificity for this discrimination was only encountered at the lowest JPEG Q value tested (Q = 5). Analysis of directly measured magnitude of image distortion from subtracted image pairs showed that the relationship between JPEG Q value and degree of image distortion underwent an upward inflection in the region of the two thresholds determined psychometrically (approximately Q = 25 to Q = 50), with 75 % of the image distortion occurring between Q = 50 and Q = 1.
It is possible to apply lossy JPEG compression to medical images without compromise of clinical image quality. Modest degrees of compression, with a JPEG Q value of 50 or higher (corresponding approximately to a compression ratio of 15:1 or less), can be applied to medical images while leaving the images indistinguishable from the original.
KeywordsMedical image compression JPEG Lossy Psychometrics Image analysis
Medical images are increasingly displayed on a range of devices connected by distributed networks, which place bandwidth constraints on image transmission. As medical imaging has transitioned to digital formats such as DICOM and archives grow in size,  optimal settings for image compression are needed to facilitate long-term mass storage requirements.
One definition of optimal medical image compression is a degree of compression that decreases file size substantially but produces a degree of image distortion that is not clinically significant. A more conservative definition of optimal image compression would require a degree of image distortion that cannot be perceived by the viewer at all. Other methods that have been used to distinguish degrees of medical image compression include pixel analysis and blinded measurements of diagnostic accuracy .
We assessed the crossover point for distortion of grayscale medical images (CT, MR, and XR modalities) by JPEG compression according to two different definitions: (1) the point at which distortion is clinically significant to the viewer and (2) the point at which any distortion can be reliably discriminated by the viewer. We additionally performed analysis of subtracted images to correlate the accumulation of increasing error pixel burden at lower JPEG Q values with the thresholds determined psychometrically.
Additional file 1: The supplemental video demonstrates the effect of continuously increasing JPEG compression from Q = 100 to Q=1 and then decreasing JPEG compression back up to Q=100 for an abdominal CT scan. The continuous range of Q values shown in the video is similar to what subjects viewed in Psychometric Experiment 1 (see Results), but in the actual experiment, the viewer was able to actively control the process of scrolling through the stack of images. (MOV 8 MB)
Image Viewing by Clinicians
Because this study aimed to determine thresholds for image distortion by JPEG compression during viewing of images in a range of clinical contexts (e.g. on a personal or clinical office computer, using a web browser, or using a portable electronic device), test images were displayed to subjects using Macintosh and Windows PCs and using both image analysis software and HTML5-compatible web browsers. Because background lux levels can impact radiological image interpretation [3, 4], background lux levels were measured using a Mastech MS8229 lux meter and maintained throughout viewing in the range of 25–100 lux.
For the presentation of continuous 100 to 1 JPEG Quality image stacks, images were presented using ImageJ64 software on a Macintosh computer with LCD screen dimensions of 1280 x 800 pixels, with images rendered at full size up to the screen resolution. Image stacks consisted of 100 images created by successively compressing an original single DICOM image into the full range of JPEG compression from JPEG Quality 100 to 1. Viewers were instructed to view the entire range of image compression from JPEG Quality 100 to 1 by scrolling through the image stack continuously using left/right arrows on the computer keyboard or scroll gestures on the computer touchpad. Viewers did not have feedback as to the degree of compression while performing this task; determinations were made solely on the basis of image appearance.
For the presentation of pairwise image comparisons, images were displayed using LCD monitors with screen resolutions of 1280 x 800 to 1280 x 1024 pixels with image presentation by way of an HTML5-compatible web browser (Google Chrome version 15) with images displayed at full size up to the screen resolution. For each pairwise comparison, viewers used the left/right arrows on the computer keyboard to rapidly switch back and forth between the two images being compared.
Clinicians in the study were practicing physicians with board certification in their primary medical specialty (Radiology, Neurology, Neurosurgery, Pulmonary/Critical Care Medicine, and Internal Medicine). A total of 8 clinicians participated in the continuous compression experiment, and a total of 10 clinicians participated in the pairwise image comparison experiment. Clinician subjects were blinded to all aspects of study design and any indicators of image compression other than intrinsic image characteristics.
Viewers assessed distortion thresholds in two different experiments: (1) determination of clinically important distortion by assessment of continuous JPEG compression from JPEG Q Value 100 to 1, and (2) determination of the level of compression that can be reliably perceived by the viewer, by assessment of a range of differently compressed image pairs.
For the continuous assessment of JPEG compression, viewers scrolled through stacks of 100 images constructed as described above with a range of JPEG compression from JPEG Q Value 100 to 1. Viewers were asked to determine the approximate point at which the image was felt to be distorted to any clinically meaningful extent, and the Q Value corresponding to this point was recorded. Viewers were allowed as much time as needed to make this determination. Each viewer assessed 40 image stacks.
For the pairwise comparison of images, viewers were shown 7 pairs of images for 8 images randomly chosen from the overall set of 40 images. For each image pair, one image was JPEG Quality = 100 and the other image was JPEG Quality = (5, 20, 35, 50, 65, 80, or 95). Each viewer was shown 7 image pairs presented in randomly chosen order and asked to determine which image of each pair (also presented in randomly chosen order) was the lower quality image. Viewers were instructed to choose an image even if they could not tell the images apart (to guess if required), and also to indicate whether they felt that their choice was a guess or not.
Random choices for image selection and order of image presentation were made with the use of a true random number generator (http://www.random.org).
For ROC plot analysis, sensitivity and specificity were calculated based on correct or incorrect identification of “image 0” or “image 1” from each image pair. Because the presentation of image pairs was chosen by random number generator, the labeling of “image 0” or “image 1” for ROC analysis was randomly chosen and the subject’s response of “image 0” or “image 1” was determined by whether the subject correctly identified the compressed image or not.
Image Pixel Difference Measurements
To determine the degree of absolute pixel differences between compressed images and a source JPEG Q Value 100 image, we performed subtraction of whole images across the range of JPEG compression from Q Value 99 to 1 using ImageJ64 software. Each successively compressed image was subtracted from the source image, yielding a stack of difference images from (Q Value 100–99) to (Q Value 100–1). Measurements were taken of the total density of difference pixels across each image in the stack, and this operation was then performed on all 40 images viewed by the subject as described above. The mean ± standard deviation total image pixel differences across the 40 images were displayed after normalization to the maximal difference in each image stack.
The conduct of this study was fully compliant with the World Medical Association (WMA) Declaration of Helsinki. Fully anonymized images without any identifying features were shown to physicians who volunteered their own time to participate. No identifying data about the individual physicians was used, stored, or transmitted as part of the study. Based on these specific study characteristics, the study was exempt from IRB review. Exempt status was confirmed by the Kaiser Foundation Research Institute IRB.
Psychometric Experiment 1: Clinically important distortion
Psychometric Experiment 2: ROC plot analysis of discrimination between compressed and original image pairs
As viewers were additionally asked in this experiment to record whether they felt that their choice was a guess, we also analyzed the relationship between JPEG Q value and the rate at which readers guessed or made the incorrect choice (Figure 3B). Consistent with the ROC plot analysis, the rate of guessing or incorrect choice rose steeply across the Q = 5 to Q = 50 range, then plateaued (Figure 3B).
Direct analysis of distortion pixels by image subtraction
Our data show that lossy JPEG compression can be applied to medical images without clinical image compromise. More subtle lossy JPEG compression (Q values of 50 or higher, roughly a compression ratio of 15:1 or less) can be applied without giving expert viewers the ability to reliably distinguish between the compressed image and the original.
The medical literature on JPEG image compression has typically presented data on compression ratios (e.g. 8:1 or 30:1). However, the software control of compression in the JPEG standard allows for direct manipulation only of Q values, not compression ratio; the compression ratio varies from image to image at a given Q value, depending on the complexity of the source image [5–7]. Since the relationship between Q value and compression ratio for a given image cannot be known a priori, it is more reasonable to present data on Q values, assuming software adherence to the standards of the Independent JPEG Group (http://www.ijg.org).
Previous work in this field has focused on relatively subtle degrees of medical image compression. For example, based on a review of the literature on compression of medical images, one group recommended a range of JPEG compression from 5:1 to 8:1. Another review of prior studies recommended this same range of compression. Similarly, consensus-based approaches have yielded estimates of acceptable compression from 5:1 to 15:1 . Another group tested higher degrees of compression following their own literature review, but they were unable to perform ROC analysis because the chosen range of compression ratios was too conservative . Of note, in the same study, JPEG compression appeared to perform better than JPEG 2000 compression at the higher levels of compression tested . This observation led us to choose JPEG compression (in contrast to JPEG 2000 compression) for our experiments.
Some work has suggested that higher degrees of compression may be acceptable. For example, one study examined the impact of JPEG 2000 compression on interpretation of mammographic digital images and found that images with compression ratios up to 60:1 were not distinguishable from source images .
Our study has limitations. We chose to focus on CT, MR, and XR modalities, all of which are grayscale, and therefore one cannot necessarily extrapolate our results to other imaging modalities, particularly color images. We also chose an approach to determine thresholds of clinically acceptable compression and the ability of readers to discriminate a compressed and original image; therefore, we did not specifically examine the ability of readers to distinguish pathology from normal anatomy, which represents a fundamentally different task.
From the data presented here and data from prior studies,[8, 9, 11–15] it is reasonable to conclude that a modest degree of JPEG compression is acceptable for many applications, particularly those involving network transmission of images.
It is possible to apply lossy JPEG compression to medical images (including CT, MR, and XR modalities) without significant compromise of clinical image quality. Regardless of whether one uses a threshold of clinically acceptable quality or a threshold of inability to distinguish the compressed image from the original, use of a JPEG Q value of 50 to 100 (an approximate compression ratio of 15:1 or lower) can be viewed as generally safe. Within the range of JPEG Q values from 50 to 100, trade-offs between quality and file size should be assessed based on the specific application or clinical need.
- Schreiber D: Swelling archives warrant closer look at compression. Radiol Manage. 2003, 25: 36-39.PubMedGoogle Scholar
- Erickson BJ: Irreversible compression of medical images. J Digit Imaging. 2002, 15: 5-14. 10.1007/s10278-002-0001-z.View ArticlePubMedPubMed CentralGoogle Scholar
- Brennan PC, McEntee M, Evanoff M, Phillips P, O’Connor WT, Manning DJ: Ambient Lighting: Effect of Illumination on Soft-Copy Viewing of Radiographs of the Wrist. American Journal of Roentgenology. 2007, 188: W177-W180. 10.2214/AJR.05.2048.View ArticlePubMedGoogle Scholar
- Chawla AS, Samei E: Ambient illumination revisited: a new adaptation-based approach for optimizing medical imaging reading environments. Med Phys. 2007, 34: 81-90. 10.1118/1.2402583.View ArticlePubMedGoogle Scholar
- Fidler A, Likar B, Skaleric U: Lossy JPEG compression: easy to compress, hard to compare. Dentomaxillofac Radiol. 2006, 35: 67-73. 10.1259/dmfr/52842661.View ArticlePubMedGoogle Scholar
- Fidler A, Skaleric U, Likar B: The impact of image information on compressibility and degradation in medical image compression. Med Phys. 2006, 33: 2832-2838. 10.1118/1.2218316.View ArticlePubMedGoogle Scholar
- Fidler A, Skaleric U, Likar B: The effect of image content on detail preservation and file size reduction in lossy compression. Dentomaxillofac Radiol. 2007, 36: 387-392. 10.1259/dmfr/66171128.View ArticlePubMedGoogle Scholar
- Braunschweig R, Kaden I, Schwarzer J, Sprengel C, Klose K: Image data compression in diagnostic imaging: international literature review and workflow recommendation. Rofo. 2009, 181: 629-636. 10.1055/s-0028-1109341.View ArticlePubMedGoogle Scholar
- Loose R, Braunschweig R, Kotter E, Mildenberger P, Simmler R, Wucherer M: Compression of digital images in radiology - results of a consensus conference. Rofo. 2009, 181: 32-37. 10.1055/s-2008-1027847.View ArticlePubMedGoogle Scholar
- Koff DA, Shulman H: An overview of digital compression of medical images: can we use lossy image compression in radiology?. Can Assoc Radiol J. 2006, 57: 211-217.PubMedGoogle Scholar
- Koff D, Bak P, Brownrigg P, Hosseinzadeh D, Khademi A, Kiss A, Lepanto L, Michalak T, Shulman H, Volkening A: Pan-Canadian evaluation of irreversible compression ratios (“lossy” compression) for development of national guidelines. J Digit Imaging. 2009, 22: 569-578. 10.1007/s10278-008-9139-7.View ArticlePubMedGoogle Scholar
- Kang BJ, Kim HS, Park CS, Choi JJ, Lee JH, Choi BG: Acceptable compression ratio of full-field digital mammography using JPEG 2000. Clin Radiol. 2011, 66: 609-613. 10.1016/j.crad.2011.02.004.View ArticlePubMedGoogle Scholar
- Kim KJ, Kim B, Lee KH, Kim TJ, Mantiuk R, Kang H-S, Kim YH: Regional difference in compression artifacts in low-dose chest CT images: effects of mathematical and perceptual factors. AJR Am J Roentgenol. 2008, 191: W30-37. 10.2214/AJR.07.3462.View ArticlePubMedGoogle Scholar
- Kim KJ, Lee KH, Kim B, Richter T, Yun ID, Lee SU, Bae KT, Shim H: JPEG2000 2D and 3D reversible compressions of thin-section chest CT images: improving compressibility by increasing data redundancy outside the body region. Radiology. 2011, 259: 271-277. 10.1148/radiol.10100722.View ArticlePubMedGoogle Scholar
- Kim TJ, Lee KH, Kim B, Kim KJ, Chun EJ, Bajpai V, Kim YH, Hahn S, Lee KW: Regional variance of visually lossless threshold in compressed chest CT images: lung versus mediastinum and chest wall. Eur J Radiol. 2009, 69: 483-488. 10.1016/j.ejrad.2007.11.035.View ArticlePubMedGoogle Scholar
- The pre-publication history for this paper can be accessed here:http://www.biomedcentral.com/1471-2342/12/24/prepub