Computer Aided Detection (CAD) for Breast MRI

Transcription

1 Technology in Cancer Research & Treatment ISSN Volume 4, Number 1, February (2005) Adenine Press (2005) Computer Aided Detection (CAD) for Breast MRI Since 1999, there has been a 40 percent increase per year in the number of breast MR studies performed in the United States. In addition, over 1200 sites in the United States have purchased surface coils for use in breast MR. This number is expected to grow to over 2,000 coils by the end of Chris Wood, M.S. Confirma, Inc. 821 Kirkland Avenue Kirkland, WA 98033, USA It is well accepted that MR sensitivity for invasive breast cancers is near 100%, but as the use of breast MRI increases, radiologists interpreting breast MR are challenged to achieve high specificity while retaining high sensitivity. Reading the large number of acquired MR images in a reasonable amount of time also becomes more important as the number of studies increases. Breast MR acquisition and image interpretation techniques have been refined through clinical optimization. The number of images to interpret, however, has increased to several hundred per case. Computer Aided Detection (CAD) algorithms have allowed radiologists to regain efficiency while maintaining optimized acquisition techniques. The first CAD system for breast MR (CADstream by Confirma, Inc.) was launched in January The CAD installed base has since grown to over 150 systems in the US. The primary reason for this quick adoption of CAD for breast MR is that the CAD software enables readers to increase their efficiency while potentially improving their overall accuracy. The full benefits CAD for Breast MR are realized when the interpreting radiologist has a thorough understanding of the algorithms used, and the limitations of CAD. Key words: CAD, Breast MR, Breast MR CAD. Introduction Indications for Breast MR Much of the demand for breast MR is driven by breast surgeons in need of accurate staging information but indications for use have recently expanded to include screening women at high-risk (1). In May 2003, the American Cancer Society released updated guidelines on breast cancer screening, recommending MRI as an effective and sensitive imaging adjunct to mammography and ultrasound (2). Breast MR Acquisition In 1986 Heywang et al. (3), showed that essentially all invasive breast malignancies enhance with after injection of gadolinium contrast. Dynamic contrastenhanced MRI studies were then described in 1989 by Kaiser (4), and it was later shown that enhancement rate and signal intensity behavior in late post-injection phases both convey diagnostically useful information (5). Corresponding Author: Chris Wood, M.S. cwood@confirma.com 49

2 50 Wood Dynamic contrast-enhanced breast MR is performed using 3D image acquisition techniques. In 60 to 90 seconds, slices covering all the breast tissue (typically both breasts) can be acquired at resolution of 1 mm or smaller. The first 3D acquisition is followed by the injection of the contrast agent, then several additional 3D acquisitions. The dynamic information is captured in less than nine minutes after the injection. Image acquisition protocols that capture the early and late post-contrast kinetics (enhancement profile) are the current standard of care. The same images may be used for morphological interpretation, or additional high resolution series may be added to the study. Breast MR Interpretation Many published studies have validated enhancement rate and washout behavior as key features that are correlated with malignancy (6). When combined with morphological interpretation, it has been shown that a sensitivity of 92% with a specificity of 92% is possible (7). The most accepted and validated list of features correlated with malignancy are captured in the BI-RADS ATLAS (8). Breast MR was included in the atlas for the first time in According to the BI-RADS committee, it is important to evaluate a finding on both morphologic and kinetic characteristics. The guidance provided by BI-RADS for interpretation of kinetics is qualitative instead of quantitative. Section F of the BI-RADS Lexicon focuses on kinetic curve assessment and reporting. The features used in the BI-RADS Lexicon to characterize curve kinetics are listed below: Initial phase (first 2 minutes) slow medium rapid Delayed phase (after 2 minutes) persistent plateau washout Figure 1 shows the guidelines for the two-phase (initial and delayed) evaluation technique described by the BI-RADS ATLAS. To interpret breast MR studies using the BI- RADS criteria, enhancing regions need to be visually identified (typically using subtractions) and delayed enhancement is typically assessed using time/intensity curves. Initial (120 seconds) fast medium slow CAD for Breast MR Delayed Note that actual percent cut-offs are not specified. Figure 1: Kinetic curve assessment. persistent plateau washout CAD has been defined as the discovery of features that may be related to malignancy (9). CAD for X-Ray mammography was the first CAD application to gain FDA approval, and like CAD for mammography, CAD for breast MR highlights features correlated with malignant breast lesions. Upon closer inspection however, at least three fundamental differences between CAD for X-Ray mammography and CAD for breast MR emerge. These differences are summarized in Table I. Given these differences, CAD for breast MR can be defined as: The automated analysis of enhancement kinetics, highlighting features related to malignancy Acquisition protocols change frequently due to innovation in pulse sequence, coil, and acquisition hardware capabilities. Differences in acquisition protocols between sites are driven by the differing capabilities of commercial MR systems and differing physician preferences. Consequently, CAD systems need to be designed to be flexible enough to adapt to protocol changes and site differences. Table I Differences between CAD for X-Ray mammography and CAD for breast MR CAD for X-Ray Features detected Morphology Contrast kinetics Method used to train computer Large training set with truth CAD for breast MR User directedspecific enhancement patterns of interest are specified by the radiologist and detected by the computer Clinical benefit(s) Increased sensitivity Reduction in interpretation time and assistance in obtaining high specificity

3 Computer Aided Detection (CAD) for Breast MRI 51 Figure 2: Typical CAD processing flow. Figure 3: Automated temporal analysis showing an enhancement pattern of slow, medium, and fast uptake. Toggle medium and fast labeling on and off Six colors indicate medium (persistent, plateau, washout) and fast (persistent, plateau, washout) Figure 4: Angiogenesis map with dual thresholds. Figure 5: Effect of temporal resolution. Curve is flattened as acquisition time increases so the threshold must be lowered. Figure 2 shows where CAD contributes to interpretation of a breast MR. Unregistered subtraction Materials and Methods Register Subtraction Figure 6: Contrasting an unregistered with a registered subtraction image. BI-RADS ATLAS-Based Interpretation of Curve Kinetics CAD systems help the physician interpret breast MR by automating extraction and interpretation of curve kinetics. Using a BI-RADS -based CAD system, curve extraction and thresholding result in angiogenesis maps, which standardize the interpretation of breast MR according to the BI- RADS ATLAS. Figure 3 shows how colors are used to map areas of enhancement and encode washout behavior. Figure 4 shows a sample angiomap with colors indicating washout behavior. Color coding provides a quick way for radiologists to find areas of significant enhancement and assess their late contrast behavior. Many factors contribute to the selection of thresholds and curve extraction. Some of these factors are described below. Contrast Dose and Injection Rate: Injection rate and dose affects contrast kinetics (10). For this reason enhancement thresholds should be adjusted depending upon the dosage and rate used. For example, a dose of 0.1 mmol/kg with

4 52 Wood thresholds of 50% and 100% (7) and a dose of 0.2 mmol/kg with corresponding thresholds of 70% and 140% (11) have both proven to result in high accuracy. Temporal and Spatial Resolution: Low spatial resolution can affect contrast kinetics for small lesions through partialvolume artifact. Lesions as small as 2 mm can enhance significantly, but slice thicknesses as high as 3 mm are still used clinically. Partial volume artifact will reduce the measured enhancement and thresholds should be lowered. As temporal resolution is reduced, rapid changes in signal enhancement can be lost due to a low sampling rate. Figure 5 shows the effect on the measured curve when reducing temporal resolution. Sampling Around Peak Contrast Concentration: Ideally, the center of k-space should be acquired coincident with the peak concentration of contrast in the tumor. If contrast concentration peaks between scans, the measured percent enhancement is reduced. Since contrast flows into a tumor faster than it washes out, even the slightest delay in sampling provides more room for error. Patient Movement: Patient movement during a volume acquisition results in blurring of the image data. This causes an artifact similar to partial volume artifacts in that the signal is spread out instead of concentrated in a few voxels. Patient movement between acquisitions causes voxel misalignment which results in inaccurate curves and the introduction of pseudo lesions. Correction for patient movement between scans is possible through non-rigid registration techniques (12). One study resulted in a reduction of pseudolesions in 92% of the cases, while not introducing any false negatives (13). Figure 6 shows the difference in subtraction images computed with, and without registration correction. Region of Interest (ROI) Size: The ROI size is an important factor to consider. If the ROI is too big, the heterogeneity of the tumor corrupts the result. If the ROI is too small, image noise may corrupt the result (14). The BI-RADS ATLAS recommends that a ROI no smaller than 3 pixels should be used. ROI Placement: Fortunately, subjective interpretation of these curves has been shown to result in agreement between observers (15). Disagreement does occur, however, when users are allowed to select their own ROI location. Manual ROI selection has resulted in up to 29% disagreement between observers (16). However, when computerized search methods are used, disagreement is eliminated and up to 38% increase in detection of washout curves is observed (17). Image Signal to Noise: Curves are displayed as percent enhancement. Parallel imaging techniques [ASSET (array spatial and sensitivity encoding technique), ipat (integrated parallel acquisition techniques), SENSE (sensitivity encoding)] can increase spatial resolution without trading off temporal resolution, but a signal-to-noise penalty needs to be considered when viewing contrast kinetics. Decreasing signal to noise may result in false positives due to random noise. Field Strength: Breast MR for diagnostic purposes is generally performed at 1.0 and 1.5T. All other factors being equal, higher field strength results in greater enhancement. Alternative Interpretation of Curve Kinetics Pharmacokinetic Modeling If one assumes that tissue and vascular spaces can be modeled as two compartments, contrast media concentration vs. time can be derived from signal intensity vs. time. From the contrast media concentration, hemodynamic parameters such as microvascular permeability and extracellular volume fraction can then be estimated. The two-compartment model assumption, however, may be invalid (18), particularly in tumors, resulting in poor fits and unreliable results (19). Concentration calculation requires measurements of the change in T1 following contrast injection, which is a time consuming exercise. Also, significant errors may be introduced because of the incorrect arterial input functions used in the model (20). While contrast kinetics are important, a local increase in vascularity and/or capillary permeability is not specific for malignant tissues: almost all benign neoplastic lesions, and many benign non-neoplastic states can exhibit variable contrast enhancement (21). Mapping pathophysiological features of breast tumors using pharmacokinetic modeling and a two compartment model is still under investigation. The validity of this model and potential clinical benefit can only be determined through continued investigation. Discussion Computer aided detection (CAD) algorithms that automate some of the interpretation tasks for the physician result in efficiency gains. It has also been shown that using CAD for breast MR may help to reduce false positives (22). The simple presence or absence of contrast enhancement, or the type of contrast enhancement, is not sufficient to either confirm or exclude the presence or absence of malignancy in breast MRI. Morphological, clinical, mammo-

5 Computer Aided Detection (CAD) for Breast MRI 53 graphic, and sonographic data should be used in conjunction with contrast kinetics to improve detection and differentiation of lesions by dynamic breast MRI. The kinetic features with the most clinical validation have been chosen by the ACR and included in the BI-RADS lexicon. These features have served as the guidance for the algorithms behind the CADstream angiogenesis maps. CADstream angiogenesis maps provide a fast and reproducible way to take images from nearly any breast MR acquisition protocol and highlight features correlated with malignancy. This results in more efficient breast MRI interpretation. Looking forward, CAD for breast MR will include additional analysis of contrast kinetics, but perhaps more importantly, quantitative analysis of morphological features. The combination of both quantitative contrast kinetics and morphology data will allow radiologists to make more evidence based decisions regarding management of suspicious lesions observed in breast tissue. When multiple breast modalities are combined, accuracy may increase even further. These improvements in CAD for breast MR will help enable custom screening protocols designed specifically for a patient s risk level. These individually optimized screening protocols have the potential to reduce mortality from breast cancer in the high-risk population. References Kriege, M., Brekelmans, C. T., Boetes, et al. Efficacy of MRI and Mammography for Breast-Cancer Screening in Women with Familial or Genetic Predisposition. N. Engl. J. Med. 351, (2004). Smith, R. A., Saslow, D., Sawyer, K. A., et al. American Cancer Society Guidelines for Breast Cancer Screening: Update. Cancer J. Clin. 53, (2003). Heywang, S. H. et al. MR Imaging of the Breast Using Gadolinium- DTPA. J. Comp. Assist. Tomogr. 10, (1986). Kaiser, W. A. and Zeitler, E. MR Imaging of the Breast: Fast Imaging Sequences With and Without Gd-DTPA: Use and Limitations. Radiology 170, (1989). Boetes, C. et al. MR Characterization of Suspicious Breast Lesions with a Gadolinium-enhanced Turbo-FLASH Subtraction Technique. Radiology 193, (1994). Orel, S. and Schnall, M.. MR Imaging of the Breast for the Detection, Diagnosis, and Staging of Breast Cancer. Radiology 220, (2001) Baum, F., Fischer, U., Vosshenrich, et al. Classification of Hypervascularized Lesions in CE MR Imaging of the Breast. Eur. Radiol. 12, (2002). Breast Imaging Reporting and Data System (BI-RADS) Breast Imaging Atlas, American College of Radiology. D Orsi, C. Computer-aided Detection: There is No Free Lunch. Radiology (2001). Joe, B. N., Bae, K. T., Chen, V. Y., et al. Dynamic MR Contrast Enhancement Characteristics of Breast Cancer: Effect of Contrast Injection Rate. Proc. of RSNA 289 (2004). Viehweg, P., Rotter, K., Laniado, M., et al. MR Imaging of the Contralateral Breast in Patients after Breast-conserving Therapy. Eur. Radiol. 14, Epub 2003 Oct 07 (2004). Kumar, R., Asmuth, J. C., Hanna, K. et al. Application of 3D Registration for Detecting Lesions in Magnetic Resonance Breast Scans. Proc. of SPiE Medical Imaging (1996). Middleton, M. S., Comstock, C. E., McNamara, M. P., et al. Quality Control Study of Image Registration in 40 Breast Cancer MR Imaging Patients using Confirma CADstream Software. Proc. of 104th ARRS 2004, 11 (2004). Mussurakis, S., Buckley, D. L., Horsman, A., et al. Dynamic MRI of Invasive Breast Cancer: Assessment of Three Region-of-Interest Analysis Methods. J. Comput. Assist. Tomogr. 21, (1997). Endo, E., Kim, S. J., Moraes, P. C., et al. Time-Signal Intensity Curve Shape Analysis on Breast MRI: Correlation with Pathology and Interobserver Variability. Proc. of 104th ARRS 2004, 61 (2004). Gabriel, H., Miller, L., Roberson, S., et al. How Reliable are Time Signal Intensity Curves in Breast MR Interpretation? Proc. of RSNA 2003, 289 (2003). Niemeyer, T., Wood, C., Stegbauer, K., and Smith, J. Comparison of Automatic Time Curve Selection Methods for Breast MR CAD. Proc. of SPIE Medical Imaging, 785 (2004). Port, R. E., Knopp, M. V., Hoffman, U,. et al. Multi-compartment Analysis of Gadolinium Chelate Kinetics: Blood-tissue Exchange in Mammary Tumors as Monitored by Dynamic MR Imaging. J. Magn. Reson. Imaging 20, (2004). Medved, M., Karczmar, G., Yang, C., et al. Semiquantitative Analysis of Dynamic Contrast Enhanced MRI in Cancer Patients: Variability and Changes in Tumor Tissue Over Time. J. Magn. Reson. Imaging 20, (2004). Parker, G. J. M. et al. Pitfalls in the Measurement of Tissue Permeability Over Short Time Scales Using a Low Temporal Resolution Blood Input Function. In: Proceedings of the 4th Annual Meeting of ISMRM, p New York (1996). Kuhl, C. K. Review Article: MRI of Breast Tumors. Eur. Radiol. 10, (2000) Lehman, C. et al. A New Method of Computer Aided Evaluation Applied to Breast MRI: Improved Specificity Without Decreased Sensitivity. Presented at Radiological Society of North America 2003 as a Hot Topic. Date Received: July 30, 2004 Date Accepted: December 10, 2004