Cassette-based Digital Mammography

Transcription

1 Technology in Cancer Research & Treatment ISSN Volume 3, Number 5, October (2004) Adenine Press (2004) Cassette-based Digital Mammography Over the past several years, digital mammography systems have been installed clinically across North America in small but growing numbers. A photostimulable phosphor-based full-field digital mammography image was evaluated in this investigation. Commonly known as computed radiography (CR), its use closely mimics the screen-film mammography paradigm. System performance using modulation transfer function (MTF) and detective quantum efficiency (DQE) metrics show MTF(2.5 mm -1 ) = 0.5, DQE(2.5 mm -1 ) = 0.3, and MTF(5.0 mm -1 ) = 0.2, DQE(5.0 mm -1 ) = 0.05, for a 26 kvp beam, 0.03 mm molybdenum tube filtration, 4.5 cm tissue attenuation, and 15 mr incident exposure to the detector. Slightly higher DQE values were measured at 32 kvp with mm rhodium tube filtration. CR mammography advantages include the ability to use existing mammography machines, where multiple rooms can be converted to digital operation, which allows overall cost savings compared to integrated digital mammography systems. Chief disadvantages include the labor-intensive handling of the cassettes prior to and after the imaging exam, lack of a direct interface to the x-ray system for recording technique parameters, and relatively slow processing time. Clinical experience in an IRB-approved research trial has suggested that digital mammography with photostimulable storage phosphors and a dedicated CR reader is a viable alternative to conventional screen-film mammography. J. Anthony Seibert, Ph.D. 1,* John M. Boone, Ph.D. 1 Virgil N. Cooper, III, Ph.D. 2 Karen K. Lindfors, M.D. 1 1 Department of Radiology UC Davis Medical Center Imaging Research Center 4701 X Street Sacramento, California USA 2 Department of Radiological Sciences UCLA Medical Center 200 UCLA Medical Plaza Los Angeles, California USA Key words: Digital mammography, Computed radiography, Digital breast imaging, CR mammography. Introduction The early detection of breast cancer by screening mammography has lead to a significant reduction in breast cancer mortality over the past decade. Dedicated x-ray equipment and screen-film detectors for mammography have incrementally advanced in performance, in part due to the implementation of strict quality control required by the Mammography Quality Standard Act (MQSA) of 1992 and subsequent revisions (1). Despite this success, there are fundamental limitations of screen-film detectors. Most significantly, the glandular breast presents an extremely large dynamic range that cannot be accommodated by the screen- * Corresponding Author: J. Anthony Seibert, Ph.D. jaseibert@ucdavis.edu Abbreviations: ACRIN: American College of Radiology Imaging Network; AEC: Automatic exposure control; CCD: Charge coupled device (camera); CR: Computed radiography; DMIST: Digital Mammography Imaging Screening Trial; DR: Direct radiography; DQE: Detective quantum efficiency; FOV: Field of view; GSV: Grayscale value; HVL: Half value layer; IP: Imaging plate; MFP: Multi-objective frequency processing; Mo: Molybdenum; MPixel: megapixel; MQSA: Mammography Quality Standards Act; mr: milli Roentgen; MTF: Modulation transfer function; NCI: National Cancer Institute; NPS: Noise power spectrum; NEQ: Noise equivalent quanta; PACS: Picture Archiving and Communications System; PMMA: Poly-methyl methacrylate (Lucite); PSL: Photostimulable luminescence; PSP: Photostimulable storage phosphor; ROI: Region of interest; SNR: Signal to noise ratio; TFT: Thin film transistor; Z: Atomic number 413

2 414 Seibert et al. film detector response. This results from the need for extremely high spatial resolution and high radiographic contrast on the film image in order to detect microcalcifications and subtle differences in x-ray attenuation between malignant lesions and surrounding glandular tissues. To achieve high contrast the film response exhibits narrow latitude that can compromise image quality in under-penetrated (glandular tissue) or over-penetrated (skin line) areas of the image. Digital mammography has received a great deal of attention over the past ten years as a possible solution to the dynamic range limitations of screen-film mammography. Significant funding from the National Institutes of Health (NIH) has focused on this area, partly due to the work of the National Digital Mammography Development Group (2). Unlike screen-film detectors, digital detectors separate the acquisition, display, and archive functions, thereby allowing each function to be independently optimized. Response of the digital detector to variations in input exposure is linear over several decades of exposure, and subsequent digital image processing can achieve contrast and spatial frequency enhancement over all incident exposure ranges typical of the thick, glandular breast, only limited by the statistics of the image formation process (the signal to noise ratio). These image processing benefits have been demonstrated with a digital detector that has considerably less spatial resolution, yet easily outperforms a screen film detector in the detection of subtle circular object contrasts of extremely small diameter (3). Despite this, several factors reduce the enthusiasm of some to adopt digital mammography, including: (i) screen-film detectors have been optimized over several decades, and are excellent for most mammography examinations; (ii) mammography demands extremely high requirements for spatial and contrast resolution, making implementation of digital detectors challenging; (iii) the high cost of large field of view digital mammography systems is hindering adoption of the technology; and (iv) many institutions still have not developed the infrastructure for electronic digital imaging capabilities, and even those who have must work in a hybrid environment (digital image displays and film displays). The current softcopy viewing environment in mammography requires the integration of specialized mammography workstations, high resolution (e.g., minimum 2,000 2,500 pixel matrix) and high luminance (> 600 cd/m 2 ) display monitors. In addition, appropriate modification of the work area to incorporate hybrid hardcopy/softcopy reading is a particularly challenging task to ensure optimal viewing conditions for both. Hanging protocols and workflow optimization are key issues to be solved, which will take some time to determine requirements and implementation strategies. Nevertheless, there is great interest in implementing digital mammography in the clinical environment. Several of the driving forces for digital technology are (i) reduced cassette handling and labor costs; (ii) higher patient throughput; (iii) image processing and direct application of computer aided diagnosis; (iv) increased reimbursement for digital screening mammography examinations, and (v) the ability to digitally archive images. Currently, the United States Food and Drug Administration has approved three full-field-of-view (FOV) digital mammography systems for clinical applications; two are based on flat-panel, thin-film-transistor (TFT) active matrix arrays, and one uses a slot-scan approach with a CCD detector array. These systems acquire an image after patient setup without the need for handling cassettes before or after the exposure. The generic term given to this class of detectors is Direct Radiography (DR). A major benefit of these systems is the prompt in-room production of an image to allow for quick analysis of positioning and image quality control. On the other hand, each of these systems has a single, fixed FOV, which can compromise the ability to image breasts that are very large or very small. All three commercially available digital mammography systems are integrated with an x-ray tube and generator, requiring extra investment in a mammography x-ray unit in addition to the detector, resulting in a high overall initial cost for a fully digital mammography practice. The main focus of this article is digital mammography performed with a cassette-based digital detector using CR technology, specifically adapted for the challenges of mammographic imaging. While CR has had great success in replacing screen-film in conventional radiography over the past decade, the application of conventional CR into mammography has been hampered by the requirements for higher spatial and contrast resolution, as well as a need for greater efficiency of image information capture. Principles of photostimulated luminescence and design of the dedicated CR mammography acquisition system are explained in Computed Radiography Prototype for Mammography of this article. An objective analysis of the CR mammography system is described in Quantitative Analysis Methods and Results, including measurements of the characteristic curve response, modulation transfer function (MTF), noise power spectrum (NPS), noise equivalent quanta (NEQ), and detective quantum efficiency (DQE) under mammographic imaging conditions. Digital processing of the raw image data, required for optimal resolution and contrast presentation is discussed, and processed digital mammography images are compared to screen-film images in Clinical Implementation of CR Mammography and Comparison to Screen-film Images. CR mammography implementation and workflow issues are considered in Workflow, Electronic Display/ Image Storage, and Quality Control, and a summary of this article is described in in Conclusions. Computed Radiography Prototype for Mammography Computed radiography is based upon photostimulable storage phosphor (PSP) detectors and photostimulated lumines-

3 Cassette-based Digital Mammography 415 cence (PSL) excitation to acquire and produce an x-ray projection image of the breast. The PSP imaging plate (IP), similar in dimensions to a conventional x-ray phosphor screen, is comprised of an unstructured layer of europium activated barium fluorobromide (BaFBr:Eu) or bromide/iodide mixture (BaFBr,I:Eu). When exposed to a variable flux of x-rays, spatially-dependent absorption of the x-ray energy generates a proportional number of electrons that are trapped in energy wells, called F centers, within the phosphor matrix; thus the term storage phosphor. Subsequent illumination of the F centers with a highly focused helium-neon or diode laser beam (630 to 680 nm wavelength red laser, energy of ev) excites the trapped electrons and elevates them to a higher energy level within the crystal. Once above a specific threshold, an electron transitions to the ground energy level, and this transition produces a photostimulated light photon of higher energy (blue light of 420 nm, energy of 3 ev) (4). The intensity of the PSL is proportional to the x-ray energy incident on the phosphor at the site of interaction. The blue light emission is captured by a light guide, amplified with a photomultiplier tube, converted into a corresponding voltage, and digitized into an integer value representing the locally absorbed x-ray intensity. The stimulating red laser light is optically filtered and thus is not recorded. This read-out of the exposed IP occurs in the CR reader, an electro-mechanical device, illustrated in Figure 1. Readout occurs as the laser actively scans the IP from left to right, is retraced to the left and begins the scan again as the plate continuously translates. The readout time is 50 seconds for the cm IP and 60 seconds for the cm IP for this specific prototype system. Pre-processing of the digital image data and Figure 1: A conventional CR reader mechanically translates the IP through an optical stage comprised of a scanning laser beam, usually of 100 µm diameter. Synchronization of the deflecting laser from a rotating mirror assembly and the translating IP designates the instantaneous row and column position (x-y position within the digital image matrix). The resulting PSL signal captured by the light guide is amplified and converted into a digital value and recorded at the specific location on the IP. An erasure module within the reader eliminates any remaining signal. transfer to the QC workstation requires another 30 and 60 seconds before the image is displayed (an elapsed time of 80 and 120 seconds for the small and large IP s, respectively). The readout involves incomplete removal of the trapped electrons within the phosphor, and thus an erasure step is required to remove residual signal prior to subsequent exposure. Depending on the erasure light intensity (typically a wide-spectrum white light is used) this step can take a considerable fraction of the total throughput time for a single IP reader. The typical erasure time is on the order of 10 to 20 seconds, and the total cycle time (insert, read, remove, and insert next IP) is approximately 60 to 80 seconds for the small and large cassettes, respectively. However, the unattenuated incident exposure on the IP when imaging thick, dense breasts requires a longer erasure time, sometimes exceeding 90 seconds, which limits throughput to less than one-half of the peak rate (e.g., 150 to 300 seconds per IP). The photomultiplier tube converts the PSL intensity captured by the light guide into a highly amplified current, which is then converted to voltage. At this stage, transformation into logarithmically amplified signals is performed either by analog circuitry or subsequent to digitization. Analog amplification reduces the dynamic range of the output signal, such that the number of bits necessary to adequately digitize the range of x-ray attenuation values is likewise reduced; 12 bit analog to digital conversion is performed on the CR mammography system. The output image is transformed to 10 bits/pixel dynamic range after the pertinent image information is identified and scaled. Initial application of CR for mammography was achieved with the conventional CR system (described above) using a low attenuation cassette housing and a high resolution detector with a thinner phosphor layer. With a laser spot diameter of 100 µm, the spatial resolution achieved was much less than that of screen-film detectors. In addition, the DQE was low (see section 2, results). In late 2000, a prototype dedicated CR mammography reader and special IP s were released in the United States, to be part of the Digital Mammography Imaging Screening Trial (DMIST) program (5). This NCI-funded ACRIN trial was designed to compare breast cancer detection performance between screen-film and digital mammography for a large number of patients (49,500 women were enrolled over a two year period). Patient accrual was completed in November 2003, and the DMIST study will continue for the next several years for analysis and comparison of the image data. Several design changes were implemented to improve the capabilities of CR specifically for mammography. The laser spot size was reduced to 50 µm, with a decrease in the corresponding sampling aperture and sampling pitch. Scanning

4 416 Seibert et al. speed of the laser beam was reduced by 50% compared to the comparable conventional reader, requiring a longer time to read out the IP. A thicker BaF(Br,I) phosphor layer (100µm) was deposited on a transparent base, and a second light guide was added (dual-side readout) to improve the collection efficiency of the PSL emitted, as shown in Figure 2. The PSL signals measured by the back and front collection guides are individually filtered by frequency bandpass functions and combined using a weighted sum, devised to maximize the noise equivalent quanta and spatial resolution (6). CR reader and workstation units in each x-ray mammography room to allow processing of exposed IPs while setting up and acquiring the next images of the mammographic series. Quantitative Analysis Methods and Results A quantitative evaluation of the prototype CR system (Fuji 5000MA, Stamford, CT) used at UC Davis for the DMIST trial was performed under mammography imaging conditions. The measurement of exposure response (characteristic curve), pre-sampled modulation transfer function (MTF p ), and normalized noise power spectra (NPS N ) were measured at 26 and 32 kvp. The noise equivalent quanta (NEQ), equal to the effective number of quanta used by the detector, and the detective quantum efficiency (DQE), an indication of the efficiency of information capture, were calculated from these data. Beam Characteristics and Computer Simulation of Spectra Figure 2: Schematic of a dual-side readout capability implemented for mammography shows two light guides and a 50 µm diameter laser beam. Scattered PSL from the laser interaction in the IP is captured on the front and back sides, and recombined with weighting algorithms applied as a function of spatial frequency to optimize the Noise Equivalent Quanta of the output image. CR mammography emulates the screen-film paradigm very closely, allowing the use of existing mammography equipment. Small (18 cm 24 cm) and large (24 cm 30 cm) FOV detectors are available (Figure 3). One CR reader can be used to process IP s from several mammography systems, and this reduces capital costs substantially. However, a significant amount of handling is required by the technologist, and the direct feedback link to evaluate positioning and image quality is not available in the room. Mainly because of throughput issues, clinical implementation of CR mammography should have as an alternate consideration individual Two mammography spectra were produced to determine characteristic responses at a relatively low effective energy (26 kvp, molybdenum (Mo) x-ray target, and 0.03 mm Mo filter), and relatively high effective energy (32 kvp, Mo x-ray target, and mm rhodium (Rh) filter). Both spectra were transmitted through a 4.0 cm block of poly-methyl methacrylate (PMMA) positioned close to the target of the mammography system, to approximate the energy-dependent attenuation of the breast in a scatter-reduced geometry. Image acquisitions were performed with the CR cassette/ip positioned on top of the cassette tunnel, without the antiscatter grid. Halfvalue-layer (HVL) of the transmitted beam through the PMMA was measured with pure aluminum filters using a manual technique with a mammography-calibrated dosimeter and ion chamber (Model #35050A dosimeter and Model # cm 3 ionization chamber, Cardinal Health, Cleveland, OH). The spectra were modeled using a previously reported technique (7) and the photon fluence (photons/mm 2 ) were estimated from these modeled spectra. Detector Characteristic Curve Figure 3: CR mammography permits the use of conventional x-ray equipment with only minor adjustments necessary for the automatic exposure control to deliver the required x-ray exposure to achieve the desired image quality in terms of signal to noise ratio (SNR). Response of the IP (note: henceforth, IP includes the imaging plate and cassette) was determined using a collimated aperture over a range of incident exposures from 4.8 mr to 39.9 mr for the 26 kvp spectrum, and from 2.3 mr to 49.3 mr for the 32 kvp spectrum. A total of 5 minutes was allowed to elapse prior to image readout to reduce variation of exposure response due to the spontaneous decay of the photostimulable phosphor. The fixed mode sensitivity number (S=121) and latitude (L = 1.82) settings of the Fuji 5000 MA reader were used for all experiments. With this setting, an incident exposure of 20 mr produced a corresponding digital value of 512 (the median value) in the 10-bit image. The latitude was the exponent to the base 10 indicating the

5 Cassette-based Digital Mammography 417 exposure range mapped symmetrically about the median value, determined by the fixed speed point. With Fuji CR systems, the sensitivity value (S number) is inversely related to the incident radiation exposure on the IP. In operation, the reader can be set to an automatic, semi-automatic, or fixed mode in terms of scaling the PSP signals into equivalent digital number values. Automatic and semi-automatic modes analyze the signals over the whole plate or a sub region of the plate, respectively. The amount of amplification (positive or negative) that is necessary to map the median value of the output signal histogram to the median digital number (511 for the 10 bit output of the Fuji system) determines the S number. Positive amplification (for low exposure) increases the S number, and negative amplification (for high exposure) decreases the S number, such that the product of the S number and the incident exposure is a constant value. When the exposure is measured in mr, this product is equal to 2400 for the Fuji 5000 MA, which means an S number of 200 under automatic or semi automatic mode indicates an average incident exposure on the IP of 12 mr. Thus, sensitivity (S number) and equivalent speed of the CR detector are related when operating in the automatic/semi-automatic mode. For the fixed mode, the sensitivity number is set, and the PSP signals corresponding to variations in incident exposure are directly mapped to a digital number without scaling. This is important for system noise and efficiency measurements. Images were transferred to an image workstation, region of interest (ROI) analysis was performed on each of the exposure patches, and a plot of the ROI mean value as a function of incident exposure was produced for each beam spectrum. As the CR system had a known logarithmic response to incident exposure, the resulting curves were fit using a least-squares algorithm: y n = a + b ln (x n ), where x n represented the mean value within the ROI for exposure n, and y n represented the corresponding mean ROI gray scale value for exposure n. All subsequent image data were transformed using floating point arithmetic to be linear with exposure for the pre-sampled MTF (MTF P ) and Noise Power Spectrum (NPS) computations. Presampled MTF An 18 cm 24 cm IP was uniformly pre-exposed to 3 mr to produce a constant background offset and to reduce noise in the low exposure regions of the line spread function (LSF) trace. Using a tightly collimated aperture arrangement using lead sheets, with the center of the region projected 2 cm from the chest wall and centered from left to right, a 10 µm tungsten slit was placed on the cassette cover and angled at approximately 3 degrees. This was performed for both scan and translation (sub-scan) readout directions of the IP, and for two beam spectra (26 and 32 kvp with PMMA phantom). An over-sampled LSF distribution was extracted using methods described by Fujita (8). Tails of the distribution were extrapolated to background. Fourier transformation of the LSF (x) produced the MTF P (f) for the scan and translation directions. A fifth-order polynomial fit to the resultant MTF curves was performed and used to interpolate the MTF P (f) to the same frequency intervals as the noise power spectrum calculation. Noise Power Spectrum Collimation was adjusted to fully expose the 18 cm 24 cm IP positioned on the cassette cover. Exposures of 4.8, 9.7, 15.4, 19.6, 30.7, and 39.3 mr for the 26 kvp spectrum and 2.3, 4.8, 9.7, 15.3, 19.7, 30.8, 38.7 and 49.3 mr for the 32 kvp spectrum were individually delivered to a freshly erased IP, with a delay of 10 minutes prior to readout. NPS estimates were computed from the linearized image data at each exposure level. The central image area closer to the chest wall side was divided into an array of pixel subimages, resulting in a frequency interval of 1/( mm) = 0.31 mm -1. A 2-D fast Fourier transform was applied to each sub-image, resulting in the frequency domain 2-D NPS (u,v), for u (column) and v (row) spatial frequencies. Individual NPS (u,v) realizations were averaged to generate an overall ensemble-averaged NPS estimate. The 1-D NPS, NPS (f), was calculated in the scan and translation directions (separately) by averaging two adjacent values along either the rows or columns, excluding the values on the axis, where f = (u 2 +v 2 ) 0.5. NPS results were made more portable by converting the fluctuations in gray scale value (GSV) to normalized fluctuations about the mean, by dividing by the square of the average large area pixel value obtained from the original images: NPS N (f) = NPS raw (f) / <GSV> 2. Noise Equivalent Quanta The NEQ (f) was determined as: NEQ (f) = MTF P 2 (f) / NPS N (f), from f = 0 mm -1 to f = f Nyquist. The MTFs specific to each x-ray spectrum were used. Detective Quantum Efficiency The DQE (f) was calculated as the ratio of the measured effective noise transfer characteristics of the detector, NEQ (f), to the estimated number of photons incident on the detector (q), as: DQE (f) = NEQ (f) / q. The value of q was determined using the modeled x-ray spectra, as discussed earlier. Each DQE curve was fit to a fourth-order polynomial over a range of f = 1.0 mm -1 to f = 8.0 mm -1 to reduce the variations in the DQE estimates. Results of Quantitative Evaluation The measured incident exposures as a function of mas and corresponding HVL s are listed in Table I. Figure 4 shows

6 418 Seibert et al. the spectra produced from the x-ray tube and transmitted through the 4.0 cm PMMA attenuator for 26 kvp, Mo target and 30 µm Mo filter and 32 kvp, Mo target and 25 µm Rh filter. Response of the detector to input exposure variations were fit to a logarithmic function with r 2 = for both input spectra as shown in Figures 5A and 5C. Image data were transformed into linear exposure space and rescaled to 12 bit dynamic range (using floating point arithmetic), with results plotted in Figures 5B and 5D. Generated spectrum Mo target/ Mo filter, 26 kvp Mo target/ Rh filter, 32 kvp Table I Beam spectra characteristics. HVL (mm Al) q (photons/mm 2 /mr) Presampled MTFs are shown in Figure 6 for the 26 kvp spectrum for the scan and translation directions. Results for the 32 kvp spectrum are within +/- 5% of these values for all spatial frequencies. Fitting coefficients for the fifth-order polynomials are indicated at the top of the graphs. An image displaying the two-dimensional NPS is shown in Figure 7. In the figure, the upper left point represents spatial frequency values u = 0 mm -1, v = 0 mm -1. Reduced noise power is noted along the u axis (scan direction) in the spectral density image beyond 8 mm -1 (dark, wide vertical band in Figure 7). The normalized noise power spectrum responses, NPS N (f), representing the noise fluctuations about the mean for a range of incident exposures are illustrated in Figure 8 for the scan and translation directions. NEQ (f) estimates representing the equivalent noise characteristics as a function of exposure are illustrated in Figure 9 for the scan and translation directions. DQE (f) estimates as a function of exposure are illustrated in Figure 10 for the scan and translation directions. The DQE (f) of an earlier model system (100 µm sampling) with single side readout is compared to the dual-side readout mammography system in Figure 11. Discussion, Quantitative Evaluation The CR mammography prototype system produces reproducible and linearizable images over an exposure range from less than 1 mr to over 50 mr, the appropriate range for mammographic applications. Signal modulation determined by the pre-sampled MTF was approximately 50% at 2.5 mm -1 and 20% at 5 mm -1 in the scan direction, and slightly lower in the translation direction. Others have shown better MTF response in the translation direction (9). Dual-side readout of the imaging plate, accuracy of the translation stage, and front /back scaling of the individual light-guide signals could be the cause Figure 4: Computer generated x-ray spectra out of the x-ray tube (A and B) and through the 4 cm PMMA block (C and D) for 26 and 32 kvp.

7 Cassette-based Digital Mammography 419 Figure 5: A) Response of the system to incident exposure variations from 1 to 50 mr at 26 kvp with a least squares logarithmic fit. B) Linearized for this difference; further investigation is warranted. MTF results for the 50 µm sampling pitch used were significantly below the ideal sinc function due to the spread of the signal distribution much larger than the effective detector aperture: MTF (5 mm -1 ) = 0.64 (sinc) versus 0.20 (CR system). Beyond frequencies of 8 mm -1 useful modulation was not existent, with consequences reflected in the NEQ/DQE measurements. and rescaled pixel value data. Fit parameters are listed in each upper right plot. C) same as A for 32 kvp. D) same as B for 32 kvp. Results of the NPS and NEQ measurements demonstrate that for the fixed mode acquisition of S=121, L=1.82, the CR system is tuned to respond over a range of 2 mr to 30 mr without significant loss of information transfer. For low exposures (below 2 mr), the DQE(f) was lower, perhaps due to the sub-optimal signal amplification for this exposure range and contributions from electronic noise sources. For high exposures (above 30 mr), structured noise, most likely Figure 6: Presampled MTF values (symbols) are shown for 26 kvp in the scan (left) and translation (right) directions, and fit with a fifth-order polynomial (solid lines) to determine response at arbitrary frequency values. Polynomial parameters are listed in the inset of each plot. Correlation (r 2 ) was >0.994 for both fits. Values near 0 mm -1 that exceeded a modulation of 1.0 were set to 1.0.

8 420 Seibert et al. due to detector and system imperfections, increased the NPS and lowered the NEQ and DQE. The highest NEQ ( 500,000 photons/mm 2 ) was reached at about 30 mr for 1 to 2 mm -1 spatial frequency range in the scan direction, and then decreased with higher incident exposure. Figure 7: The 2D NPS spectral density variations are shown using a grayscale representation. The upper left point represents u = 0 mm -1, v = 0 mm -1. Frequencies vary in the horizontal (scan) direction from u = 0 mm -1 to u = 10 mm -1 (at the midpoint) and back to u = 0.31 mm -1. Frequencies vary in the vertical direction from v = 0 mm -1 to v = 10 mm -1 (at the midpoint) back to v = 0.31 mm -1. DQE (f) was larger in the scan versus the translation direction, chiefly due to the slightly better MTF response in the scan direction, as the NPS was similar for both directions. A small, but consistent decrease in DQE with increased incident exposure occurred up to about 30 mr, and more rapidly beyond that exposure level. These results were similar to those reported by Fetterly (10), although these investigators made use of substantially different x-ray spectra (use of Al filters versus PMMA absorber). The contribution of fixed point noise with higher exposures (as visualized on the images) contributed to the noise and reduced the detective quantum efficiency. Improvements in one-dimensional shading corrections could improve the DQE values at higher exposure levels. A significant increase in DQE compared to previous CR technology was achieved due to the increased absorption efficiency of the thicker phosphor, and increased light capture with dual-side readout. Figure 8: Normalized Noise Power (mm 2 ) is logarithmically plotted versus spatial frequency as a function of incident exposure in the scan and translation directions. A) 26 kvp, Mo target/0.03 mm Mo filter, scan direction. B) 26 kvp, Mo target/0.03 mm Mo filter, translation direction. C) 32 kvp, Mo target/0.025 mm Rh filter, scan direction. D) 32 kvp, Mo target/0.025 mm Rh filter, translation direction. Decreased noise in the scan direction at spatial frequencies >8 mm -1 is noted. Above exposures of 30 mr, noise power increases, indicating the contribution from noise sources other than x-ray variations.

9 Cassette-based Digital Mammography 421 Figure 9: Noise Equivalent Quanta (photons/mm 2 ) are logarithmically plotted versus spatial frequency for the scan and translation directions as a function of incident exposure. A) 26 kvp, Mo target/0.03 mm Mo filter, scan direction. B) With higher effective energy (32 kvp, Mo target/rh filtration), the DQE was improved, possibly due to a larger number of electrons trapped in F centers relative to the total energy absorbed for the phosphor thickness. These findings suggest that higher kvp may be more efficient for CR mammography. Further investigation is warranted to determine the optimal kvp versus breast thickness for this detector system. Clinical Implementation of CR Mammography and Comparison to Screen-film Images Equipment Adjustment for Clinical CR Mammography Implementation 26 kvp, Mo target/0.03 mm Mo filter, translation direction. C) 32 kvp, Mo target/0.025 mm Rh filter, scan direction. D) 32 kvp, Mo target/0.025 mm Rh filter, translation direction. With exposures greater than 30 mr, the NEQ declines. As previously illustrated in Figure 3, the CR cassette and imaging plate are direct replacements for the screen film detectors, so physical modifications of the mammography x-ray equipment are unnecessary. X-ray absorption in the CR cassette and IP, however, are greater than with a screenfilm cassette. For the DMIST trial at UC Davis Medical Center, the phototimer system was adjusted by the AEC step controls on the generator console to deliver approximately the same dose to various thicknesses of a simulated breast phantom for both the CR, and the screen film cassettes for a gross film optical density of To achieve the results shown in Figure 12, the AEC selector switch was set to -2 when the CR detector was used for image acquisition, compared to 0 for the screen film. As each step was shown to result in an approximate 12% increase in mas, this indicates an increased attenuation of the CR cassette/ip by about 24%. At higher kvp for the greater phantom thickness, differences in the phosphor composition demonstrate a slight energy dependence in transmission through the cassettes, not unexpected due to the higher atomic number of gadolinium (Z=54) in the screen-film phosphor compared to barium (Z=34) in the CR detector for various thicknesses of simulated breast tissue of 50% glandular and 50% adipose tissue mix (BR-12).

10 422 Seibert et al. Figure 10: Detective Quantum Efficiency measurements are plotted versus spatial frequency as a function of incident exposure. A) 26 kvp, Mo target/0.03 mm Mo filter, scan direction. B) 26 kvp, translation direction. C) Figure 11: A comparison of the DQE (f) versus spatial frequency for the prototype mammography CR system with dual-side readout (top curves for 5 and 20 mr incident exposure) and the conventional CR mammography system with single-side readout (bottom curves for 5 and 21 mr incident exposure). Data was acquired at 26 kvp, and presented for the scan direction. 32 kvp, Mo target/0.025 mm Rh filter, scan direction. D) 32 kvp, translation direction. Curves represent 4 th order polynomial fits to the DQE measured data (represented as dashed curves in gray). Digital detectors have wide latitude and exposure dynamic range, allowing the image data to be acquired over a range of exposures under quantum-limited noise conditions. Depending on the desired incident exposure level to the breast and the signal to noise ratio (SNR) at the detector, adjustments of the automatic exposure control system and phototimer sensitivity must be performed. The lowest incident exposure with an SNR appropriate for accurate diagnosis is desired; in no case should the average glandular dose to the standard mammography phantom exceed 3 mgy. For the research trial at UC Davis, the goal was to achieve approximately the same incident exposure under AEC auto filter conditions for both screen film (a conventional mammography screen-film and accreditation phantom average glandular dose of about 2 mgy) and computed radiography, even though the CR detector can potentially produce a diagnostic quality image at a lower average glandular dose.

11 Cassette-based Digital Mammography 423 to a uniform flux of x-rays. Several lines of data are averaged, inverted, and normalized, creating a 1D-flat field curve that is subsequently applied to all output image data (11). This procedure reduces low and intermediate frequency noise components and improves the SNR and image quality. Of all CR system parameters measured during the two years of the DMIST trial, the shading correction calibration was most frequently required (every 6 months a shading correction was performed with the service engineer). Figure 12: Comparison of screen-film (Fuji ADM regular screens and Fuji ADM film, Fuji Photo Film, Tokyo, Japan ) and CR (dedicated dual-side mammo IP) exposures for various thicknesses of BR-12 attenuator using Automatic Exposure Control on a conventional mammography unit (LoRad M-IV, Hologic, Danbury, CT). AEC was operated under auto filter, which sets both the kvp and the filter material based upon a 100 ms pre-exposure. The system selected 25 kvp and Mo filter for the 2 and 4 cm thicknesses, 29 kvp and Mo filter for the 6 cm thickness, and 32 kvp and Rh filter for the 8 cm thickness for both screen-film (neutral or 0 position on the AEC selector) and CR ( -2 position on the AEC density selector). Image Processing The raw image data emerging from the CR reader is preprocessed, which involves two major steps. First, corrections are applied as a function of the laser beam position in the scan direction to reduce periodic spatial variations caused by changes in distribution intensity and shape of the detected PSL, and light-guide transmission efficiency variations (e.g., smudges/dirt on the light guide surface). For stationary variations in the light-guide response, a repeatable output pattern in the direction perpendicular to the scan is produced, which is along the short image axis for the 18 cm 24 cm IP, and along the long image axis for the 24 cm 30 cm IP. Shading correction algorithms reduce non-uniformities by measuring the response of an artifact-free IP exposed The second image pre-processing step is the determination of the relevant signal intensities of the breast tissues and identification of the signals in the collimated and un-collimated areas of the image. Histogram distribution analysis is the method used for determining the minimum and maximum values that correspond to the breast anatomy at the skin line and under the most highly attenuating tissues, dependent on the shape. Signals well below (underexposed, e.g., under lead disk markers) or well above (overexposed, e.g., open, un-attenuated part of the image) are designated as less important and are remapped at the extremes of the grayscale range. Thus, if the breast image is underexposed, the histogram shape is similar but is shifted to lower digital values in the raw image, and subsequent amplification rescales the data to the desired middle range of output values. Likewise, if the breast image is overexposed, the histogram distribution is shifted to the higher raw values, and subsequent de-amplification rescales the data to the middle grayscale range. The amount of increased or decreased signal amplification indicates the relative exposure of the IP that is equated to detector sensitivity or speed. Mammograms that are under or over exposed can readily be identified, based upon increased or decreased signal amplification and corresponding exposure index. Under-exposures result in increased quantum and electronic noise in the image, and reduced SNR. Over-exposures result in unacceptably higher dose to the breast and the increase in structure noise from phosphor variations that reduce SNR and compromise image quality. Figure 13: Digital mammogram image presentation. A) Unprocessed, raw image. B) Minor contrast enhancement with linear window/level adjustments. C) Moderate linear contrast enhancement, with noted loss of skin line information. D) Contrast-limited adaptive histogram equalization with slight increase in contrast. E) Same as D but with higher contrast enhancement.

12 424 Seibert et al. Figure 14A: Breast mammograms of mostly adipose tissue. Left: Screenfilm mammogram. Right: digital mammogram. Figure 14C: Breast mammograms of mostly glandular tissue. Left: Screen-film mammogram. Right: digital mammogram. simulated masses in mammograms such as contrast limited adaptive histogram equalization (14), optimum processing is critical for time-effective analysis of digital mammograms in a soft-copy environment (15). Figure 14B: Breast mammograms of heterogeneous adipose/glandular tissue Left: Screen-film mammogram. Right: digital mammogram. Post-processing the raw image is a crucial step for rendering the optimal appearance of the digital mammogram (12). This involves spatial noise reduction, edge/detail restoration, contrast enhancement and grayscale rendition processes (13) applied over the whole breast image. For mammography, specific attention to the under-penetrated (glandular tissue) and over-penetrated (skin line) areas are important. Figure 13 illustrates simple image processing applied to the raw data (Figure 13A) of a digital CR mammogram, with linear window/level adjustments (Figure 13B and C), and equalization methods using adaptive histogram equalization (Figure 13D and E). Although it has been shown that some image processing techniques do not improve the detection of For the CR system, multi-objective frequency processing (MFP) algorithms (16, 17) are currently employed, which decompose the image into various frequency ranges (image scales) to emphasize or de-emphasize edge and contrast enhancement for each scale. Reconstitution of the output image with proper weighting produces compressed dynamic range to allow a narrow-latitude grayscale rendition with good control of enhancement response in relation to spatial frequency, particularly at large transitions in signal intensity (e.g., at the skin line). A large number of control parameters are required, and considerable effort is necessary to develop effective display processing for CR mammography images. As for the comparison of screenfilm and CR mammography images illustrated in the next section, image processing parameters were preset by the manufacturer and not adjusted. Processed images were printed on dry laser film using a high-resolution printer specifically tuned for digital mammography imaging, with a maximum OD > 4.0. A single representative grayscale rendition for each image was displayed and evaluated on a conventional mammography viewer.

13 Cassette-based Digital Mammography 425 Comparison of CR and Screen-film Images Three typical screening mammogram studies are shown in Figure 14, demonstrating the similarities and the differences in the rendition of digital and analog images. Compared are the left mediolateral oblique projections of an adipose breast (Figure 14A), a heterogeneous adipose/glandular breast (Figure 14B), and a dense, glandular breast (Figure 14C). Each image was rendered from the digitization of the corresponding films using a 100 µm laser film digitizer (Lumisys 150, Kodak, Rochester, NY). The dynamic range of data presented on each film was fixed by the laser digitizer. Image comparisons reveal a visible skin line and good contrast in the adipose and glandular regions of the breast for the digital images. Although the screen-film images demonstrate good contrast, the ability to see the skin line and the local anatomy is compromised. Although not appreciated in these figures, high frequency details (i.e., microcalcification shapes) are more evident in the analog images. Robust comparisons of analog and digital mammography images, including CR images, await evaluation and follow-up by the DMIST reader study (5). Qualitative comparisons suggest that CR mammography image quality (at the same breast dose for the same technique factors) is essentially equivalent to screen-film detector systems. Workflow, Electronic Display/Image Storage, and Quality Control Workflow The dedicated CR mammography system closely mimics the screen-film paradigm, with 18 cm 24 cm and 24 cm 30 cm imaging cassettes. This allows the use of existing x-ray equipment and preserves the investment in dedicated mammography equipment. Also, multiple rooms can become digital-capable with a single, high throughput CR reader/ stacker device, making fully-digital implementation less costly. The chief disadvantages are the labor intensive handling of the imaging cassettes for acquisition and processing, with reduced throughput penalties. The cassette-based design of CR provides flexible implementation of digital mammography in a multiple room clinic with a single, high-speed external stacker reader. This central processing is similar to a daylight high-speed film processor, with the added advantage of inserting all cassettes within the reader at one time, freeing time to deal with other details of the examination. In addition, pipeline processing of the exposed IP s provides a greater throughput than a sequential read/erase single-plate reader. However, the batch-mode processing requires several minutes of downtime and less efficient productivity between studies. An alternative is a single plate CR reader for each mammography exam room that can achieve a significant increase in throughput when interfaced to the x-ray system and radiology information system (RIS) for retrieval of x-ray technique information and relevant patient demographics. After the first of 4 acquisitions, IP processing and erasure can be initiated in the room while the patient is being positioned for the next image. At the end of the fourth view, only a single IP remains for processing. In this scenario, the technologist has the ability to check positioning and image quality soon after the procedure for all four images, similar to a DR system. Single-plate readers in each room could reduce the amount of labor-intensive stack-mode processing of images, with workflow advantages similar to the fully integrated systems. On the other hand, increased costs of multiple systems versus a central unit, room utilization statistics, and workflow issues should be considered carefully. Electronic Mammographic Imaging and PACS CR digital mammography, in its current implementation, produces images with 50 µm sampling dimensions along the columns and rows of the digital matrix. For 18 cm 24 cm detectors, the matrix size is bytes, and for 24 cm 30 cm is bytes, or 33 and 55 MB of uncompressed digital data. For a typical 4-view screening CR mammogram, this represents a total of 132 to 220 MB/ study. Compression of the digital mammograms using an efficient lossless algorithm can achieve 2 to 4 reduction, dependent on the image structure, amount of breast tissue coverage, and noise characteristics of the mammogram. Even with compression, however, PACS storage and archive requirements can exceed several gigabytes/day for a busy mammography practice. In terms of display, CR digital mammograms are large, and require high-resolution display monitors. Requirements for viewing and diagnosis are similar to other digital mammography systems, including the need for high resolution monitors with a minimum of 5 Mpixel ( pixel matrix) with high luminance, and preferably larger display resolutions (e.g., 9.2 Mpixel) (18). Image presentation paradigms, specific hanging protocols for sorting and comparing mammograms, and the number of monitors that can be effectively used are crucial for radiologist efficiency. Significant demands are placed on image display hardware, software, and image processing tools. Minimum functionalities include magnification and zoom up to true size but preferably beyond, window/ level image adjustments, basic image processing for edge enhancement and tissue equalization, and advanced processing and computer aided diagnosis algorithms. Quality Control Currently, quality control procedures for digital mammography are recommended by the manufacturer and approved by

14 426 Seibert et al. the FDA. The specific digital mammography QC tests recommended by Fuji are still awaiting official FDA approvals. The physics core group of the DMIST effort, located at the Sunnybrook Health Sciences Center in Toronto, has reported on the results of the quality control program, with relevant quality control procedures for all digital mammography systems, including CR (19). Experience has been positive regarding the uptime and required calibrations of the prototype CR unit at UC Davis. Uptime was excellent over the two year period of operation, with only one-half day of unscheduled downtime due to a configuration setting error. No hardware failures were experienced over the screening mammograms and significant QC/physics studies performed on the unit. Imaging plates did not require cleaning, and images were free of artifacts for the total duration of the study. One operational issue that did require attention was the shading correction, which required recalibration at six-month intervals to reduce the effects of low frequency background signal variability. Conclusions Digital mammography is rapidly expanding into clinical use. There are several detector systems that can provide the benefits of digital acquisition, image processing, electronic display, and storage for mammography applications. A dedicated digital mammography system using photostimulable storage phosphor technology shows competitive quantitative capabilities in terms of MTF, NPS, NEQ, and DQE relative to screen-film detectors. Subjective qualitative impressions suggest comparable performance to screen-film image quality. At the time of this writing, the lower cost, ease of implementation and detector field of view flexibility are the chief advantages of CR, while the reduced throughput and extra handling of cassettes are the chief disadvantages. Consideration of single plate CR readers placed in each room in a larger mammography clinic can potentially increase throughput by processing at the same time as acquiring the next mammogram. Although with initial higher costs for multiple single-plate readers, this would allow confirmation of proper positioning and image quality prior to leaving the room, competitive with the flat-panel and slotscan digital mammography patient throughput capabilities, and allow for redundancy of image acquisition and processing should one of the readers malfunction. In summary, there is every expectation that CR will be part of the gradual evolution of digital mammography in the future. Acknowledgments This work was supported in part by the following agencies and grants: Digital Mammography Imaging Screening Trial (DMIST), ACRIN subcontract to UC Davis; California Breast Cancer Research Program Grant No. 7EB-0075; National Cancer Institute Grant No. CA-89260; National Institute for Biomedical Imaging and Bioengineering Grant No. EB ; FujiFilm Medical Systems provided the use of the prototype CR mammography unit at UC Davis. References Odle, T. G. MQSA Update. Radiol. Technol. 74, (2003). Shtern, F. Digital Mammography and Related Technologies: A Perspective from the National Cancer Institute. Radiology 183, (1992). Haus, A. G. and Yaffe, M. J. Screen-film and Digital Mammography. Image Quality and Radiation Dose Considerations. Radiol. Clin. North Am. 38, (2000). von Seggern, H., Voight, T., Knupfer, W., and Lange, G. Physical Model of Photostimulated Luminescence of X-ray Irradiated BaFBr:Eu 2+. J. Appl. Phys. 64, (1988). Pisano, E. D. and Hendrick, R. E. Digital Mammography Imaging Screening Trial. [Internet URL: (2004). Arakawa, S., Itoh, W., Kohda, K., and Suzuki, T. Novel Computed Radiography System with Improved Image Quality by Detection of Emissions from Both Sides of an Imaging Plate. Proc SPIE 3659, (1999). Boone, J. M., Fewell, T. R., and Jennings, R. J. Molybdenum, Rhodium, and Tungsten Anode Spectral Models Using Interpolating Polynomials with Application to Mammography. Med. Phys. 24, (1997). Fujita, H., Tsia, D., Itoh, T., Doi, K., Morishita, J., Ueda, K., and Ohtsuka, A. A Simple Method for Determining the Modulation Transfer Function in Digital Radiography. IEEE Trans. Medical Imaging 11, (1992). Dobbins, J. T., III, Ergun, D. L., Rutz, L., Hinshaw, D. A., Blume, H., and Clark, D. C. DQE(f) of Four Generations of Computed Radiography Acquisition Devices. Med. Phys. 22, (1995). Fetterly, K. A. and Schueler, B. A. Performance Evaluation of a Dual-side Read Dedicated Mammography Computed Radiography System. Med. Phys. 30, (2003). Seibert, J. A. Digital Radiographic Image Presentation: Preprocessing Methods. In: Advances in Digital Radiography: Categorical Course in Diagnostic Radiology Physics, pp Eds., Samei, E. and Flynn, M. J. Radiological Society of North America, Oak Brook, IL (2003). Pisano, E. D., Cole, E. B., Hemminger, B. M., Yaffe, M. J., Aylward, S. R., Maidment, A. D., Johnston, R. E., Williams, M. B., Niklason, L. T., Conant, E. F., Fajardo, L. L., Kopans, D. B., Brown, M. E., and Pizer, S. M. Image Processing Algorithms for Digital Mammography: A Pictorial Essay. Radiographics 20, (2000). Flynn, M. J. Processing Digital Radiographs of Specific Body Parts. In: Advances in Digital Radiography: Categorical Course in Diagnostic Radiology Physics, pp Eds., Samei, E. and Flynn, M. J. Radiological Society of North America, Oak Brook, IL (2003). Hemminger, B. M., Zong, S., Muller, K. E., Coffey, C. S., DeLuca, M. C., Johnston, R. E., and Pisano, E. D. Improving the Detection of Simulated Masses in Mammograms Through Two Different Imageprocessing Techniques. Acad. Radiol. 8, (2001). Pisano, E. D., Cole, E. B., Kistner, E. O., Muller, K. E., Hemminger, B. M., Brown, M. L., Johnston, R. E., Kuzmiak, C. M., Braeuning, M. P., Freimanis, R. I., Soo, M. S., Baker, J. A., and Walsh, R. Interpretation of Digital Mammograms: Comparison of Speed and Accuracy of Soft-copy Versus Printed-film Display. Radiology 223, (2002). Lure, F. Y. M., Jones, P. W., and Gaborski, R. S. Multi-resolution Unsharp Masking Technique for Mammogram Image Enhancement. Proc SPIE 2710, (1997). Ogoda, M., Hishinuma, K., Yamada, M., and Shimura, K. Unsharp