354
IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 10, NO. 2, APRIL 2006
Effects of JPEG2000 Data Compression on an Automated System for Detecting Clustered Microcalcifications in Digital Mammograms M´onica Penedo, Mar´ıa J. Lado, Pablo G. Tahoces, Associate Member, IEEE, Miguel Souto, and Juan J. Vidal
Abstract—The functionalities of the JPEG2000 standard have led to its incorporation into digital imaging and communications in medicine (DICOM), which makes this compression method available for medical systems. In this study, we evaluated the compression of mammographic images with JPEG2000 (16 : 1, 20 : 1, 40 : 1, 60.4 : 1, 80 : 1, and 106 : 1) for applications with a computer-aided detection (CAD) system for clusters of microcalcifications. Jackknife free-response receiver operating characteristic (JAFROC) analysis indicated that differences in the detection of clusters of microcalcifications were not statistically significant for uncompressed versus 16 : 1 (T = −0.7780; p = 0.4370), 20 : 1(T = 1.0361; p = 0.3007), and 40 : 1 (T = 1.6966; p = 0.0904); and statistically significant for uncompressed versus 60.4 : 1 (T = 5.8883; p < 0.008), 80 : 1 (T = 7.8414; p < 0.008), and 106 : 1 (T = 17.5034; p =< 0.008). Although there is a small difference in peak signal-to-noise ratio (PSNR) between compression ratios, the true-positive (TP) and false-positive (FP) rates, and the free-response receiver operating characteristic (FROC), figure of merit values considerably decreased from a 60 : 1 compression ratio. The performance of the CAD system is significantly reduced when using images compressed at ratios greater than 40 : 1 with JPEG2000 compared to uncompressed images. Mammographic images compressed up to 20 : 1 provide a percentage of correct detections by our CAD system similar to uncompressed images, regardless of the characteristics of the cluster. Further investigation is required to determine how JPEG2000 affects the detectability of clusters of microcalcifications as a function of their characteristics. Index Terms—Computer-aided detection (CAD), digital mammography, image compression, JPEG2000.
I. INTRODUCTION HE JPEG2000 has emerged as the new state-of-the-art standard for image compression. The recent incorporation of JPEG2000 into the digital imaging and communications in medicine (DICOM) standard facilitates its use in medical imaging applications [1]. JPEG2000 provides lossless and lossy compression, and both options have been incorporated into DICOM.
T
Manuscript received January 12, 2005; revised August 2, 2005 and September 8, 2005. M. Penedo is with the Laboratory of Medical Imaging, Unidad de Medicina y Cirug´ıa Experimental, Hospital General Universitario Gregorio Mara˜no´ n, Madrid, Spain (e-mail:
[email protected]). M. J. Lado is with the Department of Computer Science, University of Vigo, 32004 Ourense, Spain (e-mail:
[email protected]). P. G. Tahoces is with the Department of Electronics and Computer Science, University of Santiago de Compostela, 15782 Santiago de Compostela, Spain (e-mail:
[email protected]). M. Souto and J. J. Vidal are with the Department of Radiology, Complexo Hospitalario Universitario de Santiago (CHUS), University of Santiago de Compostela, 15706 Santiago de Compostela, Spain (e-mail:
[email protected]). Digital Object Identifier 10.1109/TITB.2005.864381
It is clear that the evaluation of a lossy compression method in medical imaging must consider whether the information loss is diagnostically significant for any specific clinical issue. In this study, we intend to contribute to the assessment of JPEG2000 for use in medical imaging, particularly in digital mammography. Mammography is the most effective means of detecting breast cancer before the onset of clinical symptoms. Several investigators have conducted comparative studies and have shown that digital mammography systems improve the detection of subtle lesions such as microcalcifications, the visualization of lowcontrast details or dense breast parenchymal tissue, and the definition of skin [2]–[8]. Based on these encouraging results, full-scale implementation of digital mammography technology in the near future is becoming a real possibility. However, considering that the current American Cancer Society breast cancer screening guidelines recommend annual mammography for women over 40 years of age, the increment in both transmission time and storage capacity costs, associated with the rising number of high-resolution digital mammograms per year in a hospital, would become a problem. In this scenario, an efficient data-compression scheme represents an important and viable solution to this difficulty. The JPEG2000 standard has already been evaluated in digital mammography. Sung et al. [9] evaluated JPEG2000 at several compression ratios in 20 low-resolution digitized mammograms. They concluded that there were no differences in lesion detectability with ratios of up to 20 : 1. Suryanarayanan et al. [10] investigated the effect of JPEG2000 in ten contrastdetail phantoms containing circular gold disks of different diameters (0.1–3.2 mm) and thicknesses (0.05–1.6 µm). The phantoms were acquired in a clinical full-field digital mammography system and compressed at different ratios (10, 20, and 30 : 1). Seven observers participated in the study and found no significant differences in perception up to 20 : 1, except for the disks of 1 mm in diameter. Penedo et al. [11] evaluated the effects of the JPEG2000 and the region-based method OBSPIHT (objectbased set partitioning in hierarchical trees) on the detection of clusters of microcalcifications and masses in 112 digitized mammograms at 40 and 80 : 1. Five radiologists participated in a FROC (free-response receiver operating characteristic) experiment and observed no statistical difference in detection accuracy up to 80 : 1. Digital mammography allows for utilization of advanced clinical applications, such as telemammography, offering the possibility of a second-opinion consultation from remote locations, or computer-aided detection (CAD) schemes to assist radiologists
1089-7771/$20.00 © 2006 IEEE
PENEDO et al.: EFFECTS OF JPEG2000 DATA COMPRESSION ON AN AUTOMATED SYSTEM IN DIGITAL MAMMOGRAMS
355
in the detection of radiological features that could point to different pathologies. Various full-field digital mammography systems based on different detector technologies have been approved for clinical use by the U.S. Food and Drug Administration (FDA) [12]. Accessibility of CAD technology to clinical practice involves the possibility of using high-resolution mammograms compressed with JPEG2000 standard for machine interpretation. Unlike prior studies where the assessment was based on human analysis, the aim of this study is to evaluate whether JPEG2000 can be used without significant quality degradation of the mammogram image for application to computerized detection of microcalcifications. A brief description of JPEG2000 standard, the computerized system we used, and a description of the evaluation procedure are presented in Section II. Section III presents the results of the quantitative and qualitative measurements we obtained. The paper finishes with a discussion of results and conclusions in Sections IV and V. II. MATERIAL AND METHODS A. Mammogram Database Selection A total of 450 single-view conventional mammograms were used in this study. Of the 450 mammograms, 105 were selected from the daily clinical caseload in the Department of Radiology of the Hospital Cl´ınico Universitario de Galicia (University of Santiago de Compostela, Spain), containing a total of 155 clusters of microcalcifications. The remaining 348 images were new consecutive normal cases from the screening program for women aged 50–64 years that is currently underway (since 1992) in the Galician Community. The clustered microcalcifications present in all the mammograms were confirmed by biopsy. All the images were digitized at a resolution of 2000 vertical × 2500 horizontal pixels and 4096 gray levels (12-bit precision) using a Lumiscan 85 laser scanner (Lumysis, Inc., Sunnyvale, CA). The digital images were directly stored via SCSI interface on a Sun StorEDGE A5100 magnetic disk array. A radiologist with more than ten years of experience in mammography subjectively ranked the subtlety of the clusters of microcalcifications present on the 105 mammograms on a fourlevel decision scale: 60 of the clusters were rated as obvious to detect (level 1), 44 as relatively obvious (level 2), 34 as subtle (level 3), and 17 as very subtle (level 4). The expert manually identified the size and location of each cluster of microcalcifications on a display workstation (two high-resolution 2000 × 2500 pixels and 8-bits-per-pixel monitors, both controlled through an onboard graphic card DOME, Md5 DOME Imaging Systems, Inc., Waltham, MA) by the coordinates of the center and vertexes of the smallest rectangle enclosing the cluster. These values were stored in a ground truth file and used for scoring the detection accuracy by the automated-detection procedure, as discussed in the following. Characteristics of the microcalcification clusters are shown in Fig. 1. B. The JPEG2000 Compression Method The new emerging standard for image compression, the JPEG2000 [13], is an embedded technique based on the wavelet transform that has been applied to digital mammography with
Fig. 1. Frequency of clusters included in the study according to (a) subtlety of the cluster, (b) number of microcalcifications in the cluster, and (c) size of the cluster measured as the area (in pixels) of the smallest rectangle containing the cluster (note that the entire image has 5 000 000 pixels).
promising results [9]–[11], [14]. This standard was developed to provide a set of features that the old JPEG standard did not encompass, such as control of compression ratio, progressive lossy to lossless coding/decoding, multiresolution representation, error resilience, and the possibility of encoding different regions of interest (ROI) within the image. Verification Model Software
356
IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 10, NO. 2, APRIL 2006
TABLE I AVERAGE PSNR (dB) IN THE DIGITAL IMAGES COMPRESSED WITH JPEG2000 AT DIFFERENT RATIOS
(VM8.5) 1 was used in this study to compress all mammograms with the JPEG2000 standard. Each digitized mammogram was compressed at six different compression ratios: 16 : 1, 20 : 1, 40 : 1, 60.4 : 1, 80 : 1, and 106 : 1, equivalent to a compression rate of 1.0, 0.8, 0.4, 0.25, 0.2, and 0.15 bpp (bits per pixel), respectively, and then decompressed. C. Quantitative Evaluation of JPEG2000 in Digital Mammography A parameter commonly used for measuring the degree of distortion introduced in a compression process is the peak signalto-noise ratio (PSNR), which is often measured in a logarithmic scale 2 A (1) PSNR([dB]) = 10 log10 MSE where A is the peak amplitude of the original image, and MSE is the mean square error between the original and the reconstructed image N ˆ = 1 (xn − x ˆn )2 MSE = DMSE (X − X) N n =1
(2)
where xn and x ˆn are the original and the reconstructed image pixels, respectively, and N is the number of pixels in the image. PSNR and MSE are the most common parameters used for measuring the image quality of reconstructed images after compression. The PSNR was measured in this study as an indicator of the distortion introduced in the lossy compression of the digital mammograms using the JPEG2000 standard at the six different compression rates that were tested. D. Automated Detection of Clusters of Microcalcifications All seven modalities for each digitized mammogram (1 uncompressed + 6 compressed) were processed using a CAD system for detecting clusters of microcalcifications. This scheme has been previously described and validated elsewhere [15], [16]. Briefly, the detection algorithm is a five-step process that involves: 1) detection of the breast border [17]; 2) application of one-dimensional discrete wavelet transform (DWT) within the breast region (Daubechies’ 4-tap compact support filters, fourth level decomposition), and obtainment of potential seed points of microcalcifications by reconstructing the wavelet image using only the high-frequency components of levels 3 and 4; 3) the application of a local gray level threshold to eliminate false potential microcalcifications’ seed points; 4) a 1 JPEG2000 Verification Model Software 8.5, ISO/IEC JTC1/SC29/ WG1N1894, 2000.
clustering procedure; and 5) reduction of false positives with a linear discriminant analysis. E. Analysis of Detection Accuracy The value of the discriminant function in the last step of the detection procedure was used for recording a level of confidence for the clusters of microcalcifications marked by the automated system, according to a 100-point FROC rating scale [18]. A high rating indicated high confidence that the marked location was a cluster of microcalcification. An output of the detection system was considered as a true positive (TP) if the center of the marked area was within the actual area of the cluster of microcalcifications recorded in the ground truth file. Any CAD output that did not comply with these requirements was considered as a false positive (FP). The TP and FP identified by the CAD system and their confidence ratings were collected for each image. The FROC dataset consisting of TP or FP mark-rating pairs from the 3150 observations (7 modalities × 450 images) was evaluated with the jackknife free-response receiver operating characteristic (JAFROC) analysis method 2 [19]. This method makes it possible to locate an abnormality to be accounted for in the scoring, and also allows for multiple responses and multiple lesions per image. As a figure of merit, JAFROC uses θ, which is defined as the probability that a lesion is rated higher than an FP on a normal image (FPs on abnormal images are not used in this computation). The FROC dataset is analyzed in two steps involving: 1) generation of the pseudovalues and 2) their analysis using an ANOVA (analysis of variance) model. The calculation of the pseudovalues involves removal of each image one at a time, recalculation of the figure of merit, and determination of the effect of the removal of the image on the figure of merit. A paired t test is applied internal to the ANOVA to determine differences between modalities: uncompressed versus 16 : 1, uncompressed versus 20 : 1, uncompressed versus 40 : 1, uncompressed versus 60.4 : 1, uncompressed versus 80 : 1, and uncompressed versus 106 : 1. A Bonferroni adjustment was used for the six pairwise comparisons. If the observed p value was less than 0.008 (p < 0.05/6), the null hypothesis of no differences between the modalities was rejected. III. RESULTS Table I summarizes the distortion introduced in the lossy compression of digital mammograms at six different ratios with JPEG2000. As expected, the PSNR decreases as the compression ratios increase. However, there is a small difference in 2 The JAFROC software used in this study (version 1.03) is available for download from the website http://jafroc.radiology.pitt.edu.
PENEDO et al.: EFFECTS OF JPEG2000 DATA COMPRESSION ON AN AUTOMATED SYSTEM IN DIGITAL MAMMOGRAMS
357
TABLE II PERFORMANCE OF THE CAD SYSTEM IN AUTOMATIC DETECTION OF CLUSTERS OF MICROCALCIFICATIONS USING IMAGES COMPRESSED WITH JPEG2000 AT DIFFERENT RATIOS AND UNCOMPRESSED (1 : 1) IMAGES
TABLE III PERCENTAGE OF CORRECT DETECTIONS OF THE CAD SYSTEM WITH ORIGINAL AND COMPRESSED IMAGES RELATED TO THE CHARACTERISTICS OF THE CLUSTER OF MICROCALCIFICATIONS
TABLE IV AVERAGE VALUES OF THE FROC FIGURE OF MERIT (θ) QUANTIFYING CAD PERFORMANCE IN THE DETECTION OF CLUSTERS OF MICROCALCIFICATIONS FOR ALL COMPRESSION RATIOS USED IN THE STUDY
PSNR between compression ratios. The TP and FP findings obtained with the computerized system for compressed mammograms are shown in Table II. Note the substantial difference in TP and FP values for compression ratios greater than 40 : 1. Table III summarizes how the subtlety, size, and number of microcalcifications within the cluster affect the sensitivity of the CAD system when using compressed images. The automated scheme obtains the same percentage of correct detections with uncompressed images and images compressed at 16 : 1, regardless of the characteristics of the clusters. A very slight difference in the CAD performance is found between original images and images compressed at 20 : 1. However, there is a remarkable decrease in the sensitivity of the automated-detection system when the images used are compressed at 60.4 : 1, 80 : 1, and 106 : 1. The CAD sensitivity results for images compressed at 40 : 1 show more variability compared to uncompressed images depending on the cluster characteristics. For example, we find similar performance as the size of the cluster increases, when the cluster presents 10–20 microcalcifications or with clusters categorized as “very subtle,” and a decrease in sensitivity for all other situations.
The results of the JAFROC analysis of the collected data are shown in Table IV. JAFROC analysis indicated that differences in the detection of clusters of microcalcifications were not statistically significant for uncompressed versus 16 : 1 (T = −0.7780; p = 0.4370), uncompressed versus 20 : 1 (T = 1.0361; p = 0.3007), uncompressed versus 40 : 1 (T = 1.6966; p = 0.0904); and statistically significant for uncompressed versus 60.4 : 1 (T = 5.8883; p < 0.008), uncompressed versus 80 : 1 (T = 7.8414; p < 0.008), and uncompressed versus 106 : 1 (T = 17.5034; p =< 0.008). These results are consistent with the results shown in Tables II and III, where we found a considerable difference in TP and FP values and percentage of sensitivity at a compression ratio greater than 40 : 1. Figs. 2 and 3 show examples of ROIs compressed with JPEG2000 at different ratios and the different outputs obtained with the CAD algorithm. IV. DISCUSSION Although some controversy still exists about the use of lossy compression in medical imaging, the FDA allows medical image
358
IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 10, NO. 2, APRIL 2006
Fig. 2. Magnified region of interest containing a subtle cluster of microcalcifications. The performance of the CAD system in identifying this region was (a) TP region at 16 : 1, (b) TP region at 20 : 1, (c) TP at 40 : 1, (d) false-negative region at 60.4 : 1, (e) false-negative region at 80 : 1, and (f) false-negative region at 106 : 1. At high compression ratios, JPEG2000 significantly affects the performance of the CAD system in the detection of clusters of microcalcifications.
management devices to use irreversible compression if it is certain that the information loss will not affect diagnosis. This has motivated research studies to evaluate compression in medical imaging, since the particular compression ratio that would be clinically acceptable depends on the radiological features of the medical imaging modality and the compression technique used [20]. In particular, the importance of digital mammography in the screening and diagnosis of breast cancer requires extensive evaluation of any compression method that would be applied to this image modality, since small changes in the performance obtained with compressed mammograms would affect a considerable number of individuals if the screening population is large. The interest of the evaluation of the JPEG2000 standard for medical imaging over other compression methods resides in its acceptance as an image compression option by the DICOM
Fig. 3. Magnified region of interest of a normal case. The performance of the CAD system for this region was (a) true-negative region at 16 : 1, (b) truenegative region at 20 : 1, (c) true-negative region at 40 : 1, (d) FP region at 60.4 : 1, (e) FP region at 80 : 1, and (f) FP region at 106 : 1. The “rice artifacts” introduced with JPEG2000 significantly affect CAD performance at ratios above 60.4 : 1.
standard. JPEG2000 can be used in medical systems such as picture archiving and communications system (PACS) and telemammography to reduce storage and transmission costs, facilitating digital mammography as a medical imaging modality. Other compression methods, such as JPEG, JPEG-LS, and runlength encoding (RLE), are also accepted in DICOM, but the wavelet-based JPEG2000 method has emerged as a better technology for image compression. One of the main new features of JPEG2000 for medical applications is the possibility of ROI encoding. Digital mammography would benefit from the possibility of compressing regions with radiological signs of breast cancer at a quality different from the rest of the image. Penedo et al. evaluated an adaptation of the state-of-the-art compression set partitioning in hierarchical trees (SPIHT) method [21] to compress only the breast region in a digital mammogram, called object-based SPIHT (OBSPIHT) [11], [14]. Contrary to the ROI encoding option of JPEG2000, OBSPIHT applies the wavelet transform independently to each region. More efficient coding of the ROI is expected with OBSPIHT, but at the
PENEDO et al.: EFFECTS OF JPEG2000 DATA COMPRESSION ON AN AUTOMATED SYSTEM IN DIGITAL MAMMOGRAMS
expense of higher computational costs compared to the wavelet transform applied to the entire image as in JPEG2000. Although OBSPIHT showed results competitive with JPEG2000 in the observer performance study with radiologists, the use of the standard ensures that compression capabilities can be integrated into medical imaging systems. The FDA approval of various commercially available CAD systems for digital mammography has directed our attention to validation of the JPEG2000 standard for using computer-aided programs in clinical practice. Parameters such as PSNR or MSE for measuring distortion in compressed images could be of limited value for medical images. However, the FDA requires measurement of the normalized mean square error for each compression factor used by medical management devices using lossy compression methods [22]. As in the study by Sung et al. [9], we measured the PSNR of digitized mammographic images compressed at different ratios with JPEG2000, with small differences in PSNR when the compression ratio increases. Our study represents the first evaluation of the performance of a computerized algorithm dedicated to the detection of clusters of microcalcifications when using high-resolution mammograms compressed with the JPEG2000 standard. We performed a FROC experiment with one observer (the CAD system) and included a set of 450 digitized mammographic images. Other research papers have used radiologists in observer performance studies to evaluate the impact of JPEG2000 compression on digital mammography [9]–[11]. For example, Sung et al. [9] performed an ROC (receiver operating characteristic) [23], [24] experiment with three radiologists and a set of 20 low-resolution digitized mammograms. The ROC methodology is applicable only to tasks that call for a binary decision from the observer, i.e., presence or absence of a lesion, regardless of its location [19]. They reported statistical differences above 15 : 1 in quality for lesion detectability of the compressed mammograms (using an image rating scale of certainly acceptable to certainly not acceptable for diagnosis). As a difference from their study, the FROC methodology that we used allows the observer to report multiple abnormalities and consider the correct location of a lesion, which is more similar to the normal clinical diagnostic task in mammography. Furthermore, it has been shown that the JAFROC analysis we used in our study outperforms the ROC method in statistical power [19] since JAFROC generalizes to both population of images and readers. Although their statistical analysis considered only quality for diagnosis and not detection ability, a comparison with their results leads to an assumption that an expert radiologist perceives significant differences on JPEG2000 compressed mammographic images at lower compression ratios than a computer scheme. However, detection accuracy of microcalcifications is significantly reduced when using low pixel size and pixel depth images [25], [26]. Accordingly, their results could reflect the effect of JPEG2000 compression but could also be the consequence of including a low-resolution image dataset in their study. In [11], JPEG2000 was evaluated in a FROC observer performance study with five radiologists. They found, using JAFROC analysis, that compression at 80 : 1 with JPEG2000 can be used without statistically decreasing
359
the diagnostic accuracy for detecting clusters and masses in high resolution digitized mammograms. In our study, we found that the detection ability of the CAD system significantly decreases from 60.4 : 1. We deduce that the detection of clusters of microcalcifications is more susceptible to JPEG2000 compression when using a computer-aided program than when the detection is performed by a human observer. In another evaluation of JPEG2000, Suryanarayanan et al. acquired ten images with a clinical full-field digital mammography system of a phantom containing disks of different diameters and thicknesses. They studied the effect of compression on contrast–detail characteristics of digital mammography using seven human observers. Their results showed that compression up to 20 : 1 does not have a statistically significant impact on contrast–detail characteristics for disks between 0.63–0.8 mm in diameter. As is pointed out in their paper [10], the results are for images of the type they investigated. Although it is difficult to compare our work with their results based on phantom images, it appears that JPEG2000 compression affects the ability to detect small low-contrast objects (such as microcalcifications), but with respect to the detection of clusters of microcalcifications (clustered microcalcifications are more clinically significant than isolated microcalcifications [27]), no significance difference is observed at higher compression ratios. The subtlety of the clusters in this study was determined by only one expert radiologist. The images also contain a considerable number of “obvious” and “relatively obvious” lesions. It could be pointed out that the automated scheme easily detects these two categorized types of clusters and that the categorization used may not be objective enough. While these are all legitimate concerns, as long as this possible bias applies equally to the different compressed images used in the automated method, these factors are not expected to affect the conclusion of our study. In this study, the aim was to assess how the JPEG2000 compression affects the results of a CAD scheme, not to evaluate the performance of the automated system. The results in Table III show that the same percentage of “very subtle” clusters was detected by the computerized system when using original images or images compressed up to 40 : 1. There are other important aspects that deserve comment. If we analyze the results shown in Table III, we find that we can establish two different levels of the CAD technology performance, depending on the cluster size. When the detection algorithm is applied to images compressed at 40 : 1, 60.4 : 1, 80 : 1, and 106 : 1, the sensitivity rises as the cluster size increases. The opposite situation can be found for uncompressed images and compression ratios of 16 : 1 and 20 : 1. In these cases, sensitivity increases as the cluster size decreases. This may be due to the fact that when the compression ratio increases, a great quantity of detailed information is missed. Thus, as the cluster diminishes in size, the possibility of missing this information increases. The results of this study suggest that digital mammograms compressed up to 40 : 1 with JPEG2000 can be used by our CAD system without significant loss in detection performance compared to uncompressed images. However, further investigation is required to determine how JPEG2000 affects the detectability of clusters of microcalcifications as a function of their characteristics.
360
IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 10, NO. 2, APRIL 2006
Our study is limited somewhat by the fact that we used an automated procedure for detection of clustered microcalcifications. The automatic detection of other important radiological signs in digital mammography, such as masses, should be also evaluated when using images compressed with JPEG2000. V. CONCLUSION The JPEG2000 standard has been included as part of the DICOM standard, facilitating its use in applications such as automated procedures for the detection of lesions. Our study is the first to evaluate the effects of JPEG2000 compression with a CAD digital mammography system dedicated to the detection of clustered microcalcifications. The performance of the CAD system is not significantly reduced when using images compressed up to 40 : 1 with JPEG2000 compared to uncompressed images. Characteristics of the clustered microcalcifications, such as subtlety, size, and number of microcalcifications within the cluster, play an important role in the quality of mammographic images compressed with JPEG2000. With our CAD system, mammographic images compressed up to 20 : 1 provide a percentage of correct detections similar to uncompressed images, regardless of the characteristics of the cluster. ACKNOWLEDGMENT The authors are grateful to Dr. D. Chakraborty, Associate Professor of Radiology and Bioengineering, Department of Radiology, University of Pittsburgh, for his help and advice in the statistical analysis. REFERENCES [1] D. A. Clunie, “Lossless compression of grayscale medical images— Effectiveness of traditional and state of the art approaches,” in Proc. Medical Imaging 2000: Pacs Design and Evaluation—Engineering and Clinical Issues, vol. 1; Proc. Society of Photo-Optical Instrumentation Engineers (SPIE), 2000, pp. 74–84 [2] O. Jarlman, L. Samuelsson, and M. Braw, “Digital luminescence mammography. Early clinical experience,” Acta Radiol., vol. 32, pp. 110–113, Mar. 1991. [3] D. S. Brettle, S. C. Ward, G. J. Parkin, A. R. Cowen, and H. J. Sumsion, “A clinical comparison between conventional and digital mammography utilizing computed radiography,” Br. J. Radiol., vol. 67, pp. 464–468, May 1994. [4] A. R. Cowen, G. J. S. Parkin, and P. Hawkridge, “Direct digital mammography image acquisition,” Eur. Radiol., vol. 7, pp. 918–930, Aug. 1997. [5] L. A. Venta, R. E. Hendrick, Y. T. Adler, P. DeLeon, P. M. Mengoni, A. M. Scharl, C. E. Comstock, L. Hansen, N. Kay, A. Coveler, and G. Cutter, “Rates and causes of disagreement in interpretation of fullfield digital mammography and film-screen mammography in a diagnostic setting,” AJR Amer. J. Roentgenol., vol. 176, pp. 1241–1248, May 2001. [6] E. A. Berns, R. E. Hendrick, and G. R. Cutter, “Performance comparison of full-field digital mammography to screen-film mammography in clinical practice,” Med. Phys., vol. 29, pp. 830–834, May 2002. [7] S. Suryanarayanan, A. Karellas, S. Vedantham, H. Ved, S. P. Baker, and C. J. D’Orsi, “Flat-panel digital mammography system: Contrast-detail comparison between screen-film radiographs and hard-copy images,” Radiology, vol. 225, pp. 801–807, Dec. 2002. [8] P. Skaane and A. Skjennald, “Screen-film mammography versus fullfield digital mammography with soft-copy reading: Randomized trial in a population-based screening program—The Oslo II study,” Radiology, vol. 232, pp. 197–204, Jul. 2004. [9] M.-M. Sung, H.-J. Kim, E.-K. Kim, J.-Y. Kwak, J.-K. Yoo, and H.-S. Yoo, “Clinical evaluation of JPEG2000 compression for digital mammography,” IEEE Trans. Nucl. Sci., vol. 49, no. 3, pp. 827–832, Jun. 2002.
[10] S. Suryanarayanan, A. Karellas, S. Vedantham, S. M. Waldrop, and C. J. D’Orsi, “A perceptual evaluation of JPEG 2000 image compression for digital mammography: Contrast-detail characteristics,” J. Digit. Imag., vol. 17, pp. 64–70, Mar. 2004. [11] M. Penedo, M. Souto, P. G. Tahoces, J. M. Carreira, J. Villal´on, G. Porto, C. Seoane, J. J. Vidal, K. S. Berbaum, D. P. Chakraborty, and L. L. Fajardo, “FROC evaluation of JPEG2000 and object-based SPIHT lossy compression on digitized mammograms,” Radiology, vol. 237, pp. 450–457, 2005. [12] J. J. James, “The current status of digital mammography,” Clin. Radiol., vol. 59, pp. 1–10, Jan. 2004. [13] M. Rabbani and R. Joshi, “An overview of the JPEG2000 still image compression standard,” Signal Process: Image Commun., vol. 17, pp. 3– 48, 2002. [14] M. Penedo, W. A. Pearlman, P. G. Tahoces, M. Souto, and J. J. Vidal, “Region-based wavelet coding methods for digital mammography,” IEEE Trans. Med. Imag., vol. 22, no. 10, pp. 1288–1296, Oct. 2003. [15] M. J. Lado, P. G. Tahoces, A. J. Mendez, M. Souto, and J. J. Vidal, “A wavelet-based algorithm for detecting clustered microcalcifications in digital mammograms,” Med. Phys., vol. 26, pp. 1294–1305, Jul. 1999. [16] M. J. Lado, P. G. Tahoces, A. J. Mendez, M. Souto, and J. J. Vidal, “Evaluation of an automated wavelet-based system dedicated to the detection of clustered microcalcifications in digital mammograms,” Med. Inform. Internet Med., vol. 26, pp. 149–163, Jul./Sep. 2001. [17] A. J. Mendez, P. G. Tahoces, M. J. Lado, M. Souto, J. L. Correa, and J. J. Vidal, “Automatic detection of breast border and nipple in digital mammograms,” Comput. Methods Programs Biomed., vol. 49, pp. 253– 262, May 1996. [18] D. Chakraborty and L. Winter, “Free-response methodology: Alternate analysis and a new observer-performance experiment,” Radiology, vol. 174, pp. 873–881, Mar. 1990. [19] D. P. Chakraborty and K. S. Berbaum, “Observer studies involving detection and localization: Modeling, analysis, and validation,” Med. Phys., vol. 31, pp. 2313–2330, Aug. 2004. [20] B. Erickson, A. Manduca, P. Palisson, K. Persons, F. Earnest, IV, V. Savcenko, and N. Hangiandreou, “Wavelet compression of medical images,” Radiology, vol. 206, pp. 599–607, Mar. 1998. [21] A. Said and W. Pearlman, “A new fast and efficient image codec based on set partitioning in hierarchical trees,” IEEE Trans. Circuits Syst. Video Technol., vol. 6, no. 3, pp. 243–250, Jun. 1996. [22] FDA, “Guidance for the submission of premarket notifications for medical image management devices,” Center for Devices and Radiological Health of the Food and Drug Agency. U.S. Dept. Health and Human Services, Washington, DC, 2000. [23] C. E. Metz, “Basic principles of ROC analysis,” Semin. Nucl. Med., vol. 8, pp. 283–298, 1978. [24] C. Metz, “Evaluation of digital mammography by ROC analysis,” in Proc. Digital Mammography, M. L. G. K. Doi, R. M. Nishikawa, and R. A. Schmidt, Eds., ser. Experta Medica Int. Congress, Amsterdam, The Netherlands: Elsevier, 1996, pp. 61–68. [25] H. P. Chan, C. J. Vyborny, H. MacMahon, C. E. Metz, K. Doi, and E. A. Sickles, “Digital mammography—ROC studies of the effects of pixel size and unsharp-mask filtering on the detection of subtle microcalcifications,” Invest. Radiol., vol. 22, pp. 581–589, Jul. 1987. [26] H. P. Chan, L. T. Niklason, D. M. Ikeda, K. L. Lam, and D. D. Adler, “Digitization requirements in mammography—Effects on computer-aided detection of microcalcifications,” Med. Phys., vol. 21, pp. 1203–1211, Jul. 1994. [27] R. M. Nishikawa, M. L. Giger, K. Doi, C. J. Vyborny, and R. A. Schmidt, “Computer-aided detection of clustered microcalcifications—An improved method for grouping detected signals,” Med. Phys., vol. 20, pp. 1661–1666, Nov./Dec. 1993.
M´onica Penedo was born in Barcelona, Spain, on November 11, 1973. She received the B.Sc. and Ph.D. degrees in physics from the University of Santiago de Compostela, Santiago de Compostela, Spain, in 1997 and 2002, respectively. She is presently a Senior Engineering Researcher at the Medical Imaging Laboratory at the Hospital General Universitario Gregorio Mara˜no´ n, Madrid, Spain. Her research interests include medical image compression in breast imaging, as well as waveletbased statistical methods for neuroimage studies.
PENEDO et al.: EFFECTS OF JPEG2000 DATA COMPRESSION ON AN AUTOMATED SYSTEM IN DIGITAL MAMMOGRAMS
Mar´ıa J. Lado was born in Santiago de Compostela, Spain, on June 6, 1971. She received the B.Sc. and Ph.D. degrees in physics from the University of Santiago de Compostela, Santiago de Compostela, Spain, in 1995 and 1999, respectively. She is presently a Professor of Computer Science at the University of Vigo, Ourense, Spain. Her research intrests include computer-aided diagnosis in breast and chest/CT imaging, as well as computeraided systems for teaching purposes.
Pablo G. Tahoces (A’90) was born in Ponferrada, Spain, on July 17, 1965. He received the B.Sc. and Ph.D. degrees in physics from the University of Santiago de Compostela, Santiago de Compostela, Spain, in 1988 and 1992, respectively. He is presently an Associate Professor of Computer Science at the University of Santiago de Compostela. His research interests include computeraided diagnosis in breast and chest/CT imaging, image compression, as well as medical imaging systems. Dr. Tahoces is a member of the IEEE Engineering in Medicine and Biology Society and of the American Association of Physicists in Medicine.
361
Miguel Souto was born in Caracas, Venezuela, on May 31, 1959. He is a medical doctor and received the Ph.D. degree in 1990 from the University of Santiago de Compostela, Santiago de Compostela, Spain. He is presently a Professor of Radiology at the University of Santiago de Compostela. His research interests include computer-aided diagnosis in breast and chest/CT imaging, image compression, as well as medical imaging systems.
Juan J. Vidal was born in Santiago de Compostela, Spain, on January 1, 1941. He is a medical doctor and received the Ph.D. degree in 1975 from the University of Santiago de Compostela, Santiago de Compostela, Spain. He is presently a Professor of Radiology at the University of Santiago de Compostela. His research interests include computer-aided diagnosis in breast and chest/CT imaging.
