C HAPTER 22 C OMPUTER -A IDED D IAGNOSIS

OF

B REAST C ANCER ON MR I MAGING

Qiu Wu Department of Electrical and Computer Engineering, The University of Texas at Austin Mia K. Markey Department of Biomedical Engineering, The University of Texas at Austin

C ONTENTS 22.1 Medical Imaging for Breast Cancer Detection and Diagnosis / 740 22.2 Magnetic Resonance Imaging of the Breast / 740 22.3 Dynamic Contrast-Enhanced Breast MRI / 741 22.4 Computer-Aided Detection and Diagnosis / 744 22.5 Developing CADe/CADx for DCE Breast MRI / 745 22.5.1 Image registration / 745 22.5.2 Lesion localization / 747 22.5.3 Lesion segmentation / 748 22.5.4 Feature extraction / 749 22.5.5 Feature selection / 750 22.5.6 Lesion classification / 751 22.5.7 CADe/CADx evaluation and validation / 751 22.6 Future Directions / 753 Acknowledgments / 754 References / 755

739

740

22.1 M EDICAL I MAGING D IAGNOSIS

C HAPTER 22

FOR

B REAST C ANCER D ETECTION

AND

Breast cancer is the second leading cause of cancer death for American women today.1 The key to increasing the survival rate is early detection and treatment. Medical imaging is essential to breast cancer screening and diagnosis. Currently, x-ray mammography is the primary screening modality for breast cancer.2 Mammography is a radiographic examination technique that uses x rays to image breast tissue to detect breast pathology. Denser tissue absorbs more x rays and results in a bright region on the image.3 Mammography attempts to identify structural or morphological differences that can indicate presence of cancer, such as masses, microcalcifications, and architectural distortions.4 Traditionally, the image has been recorded on film. Recently, digital mammography techniques have been adopted that record the image in digital format.5 Mammography is a well-developed technology that offers high-quality images at low radiation doses for the majority of patients. Unfortunately, 10–30% breast cancers are not detected on mammography6–8 and the positive predictive value (PPV) of mammography is less than 35%.9 Consequently, other imaging modalities are used in conjunction with x-ray mammography for detection and diagnosis. Moreover, following the diagnosis of cancer, medical imaging is used for treatment planning, monitoring treatment progress, and surveillance for disease recurrence. The most widely used adjunctive modality for breast imaging is ultrasound. Ultrasound is routinely used to further evaluate suspicious abnormalities identified on screening mammography or a clinical exam. Ultrasound is particularly valuable for distinguishing between cysts and solid lesions and for examining younger women with dense breasts.4 The roles of other imaging modalities in breast cancer detection and diagnosis are rapidly evolving. Tomosynthesis,10 newer ultrasound methods such as 3D ultrasound,11 nuclear medicine methods such as sestamibi breast scintigraphy and positron emission tomography (PET),12,13 and magnetic resonance imaging (MRI) are all exciting areas of development in breast cancer care. Each of these alternatives are being sought in order to overcome at least one of the two major inherent limitations of x-ray mammography: the information loss of visualizing a 3D structure in 2D projections and the lack of functional insight regarding the biological processes of the breast tissue imaged.

22.2 M AGNETIC R ESONANCE I MAGING

OF THE

B REAST

MRI primarily images the nuclear magnetic resonance (NMR) signal from the hydrogen nuclei of the tissue.14 The recovery of longitudinal magnetization under the external magnetic field is characterized with a time constant T1 , the longitudinal relaxation time. Similarly, the recovery of transverse magnetization is characterized with a time constant T2 , the transverse relaxation time. The recovery of transverse

C OMPUTER -A IDED D IAGNOSIS

OF

B REAST C ANCER

ON

MR I MAGING

741

magnetization in an inhomogeneous magnetic field is determined by the time constant T2∗ . When a specially designed sequence is used, T1 , T2 , or T2∗ signals from the tissue can be imaged. Usually T1 and T2 are unique biophysical characteristics of the tissue and can thus be used to provide contrast between different tissues on a T1 /T2 /T2∗ -weighted image. By applying a 3D encoded magnetic field, MRI enables a truly 3D examination of breast tissues [Fig. 22.1(a)], but even 3D structural information is insufficient to reliably distinguish between abnormal and normal tissues. The need for functional data is the driving force behind the development of dynamic contrast-enhanced MRI.

22.3 DYNAMIC C ONTRAST-E NHANCED B REAST MRI Functional imaging is required in order to recognize the distinctions between tumor cells and normal cells that exist at the molecular level, such as differences in cellular composition, permeability, and microvessel density.15 The use of contrast agents in MRI enables the visualization of functional changes, particularly angiogenesis, when sequential MRI scans are acquired [Fig. 22.1(b)].16,17 Dynamic contrast-enhanced (DCE) breast MRI, in which the breast is imaged before, during, and after the administration of a contrast agent, provides a noninvasive assessment of the microcirculatory characteristics of tissues in addition to traditional anatomical structure information. In a DCE MRI exam, a contrast agent such as Gadolinium diethyltriaminepentaacetic acid (Gd-DTPA) diffuses into the extravascular extracellular space (EES) via the capillaries and accumulates in tissues with high vascularity, and subsequently leaks back into the vascular space and is eventually excreted from the body.15 The diffusion process is governed by the kinetic properties of the target tissues. The concentration of the contrast agent alters the relaxation time of water protons in the surrounding tissue; thus, the accumulated amount of contrast agent around the targeted tissue is reflected in the MR image intensity. Since there is contrast agent uptake and washout over time, dynamic images are produced during sequential MRI scans. In DCE MRI, pulse sequences can be chosen so that the resulting images are selectively sensitive to different vascular and kinetic characteristics. With appropriate use of contrast agents, dynamic T1 /T2∗ -weighted images can be acquired. The net effects on the image intensities of targeted tissues on the resulting T1 -weighted and T2∗ -weighted images are opposite.18,19 On T1 weighted images, there is “enhancement,” while on T2∗ -weighted images there is “darkening.”18–20 However, the term dynamic contrast enhancement is used in either case. T1 -weighted images have been most commonly used in assessing breast tissue. Unless otherwise stated, the DCE breast MR images discussed in this chapter are T1 -weighted images. The dynamic image intensities reflect the physiological nature of the targeted tissue. For any location in the image, the dynamic image intensities can be viewed as a multidimensional signal, usually referred to as the kinetic enhancement curve [Fig. 22.1(c)]. Previous studies have shown that the kinetic enhancement curves of

742

C HAPTER 22

(a)

(b)

(c) Figure 22.1 DCE breast MRI produces 4D data—three spatial dimensions plus time. (a) Axial, coronal, and sagittal planes of a 3D breast image. (b) Axial view of precontrast and first two postcontrast images. The enhancement can be observed in the postcontrast images on the left breast. (c) Shows a kinetic enhancement curve of a voxel.

C OMPUTER -A IDED D IAGNOSIS

OF

B REAST C ANCER

ON

MR I MAGING

743

malignant and benign lesions are different, on average (Fig. 22.4). Thus, in addition to the morphological features of the lesion, kinetic features are recommended for use in determining the pathological nature of the lesion.21 In this regard, the use of contrast agents has increased both the sensitivity and specificity of breast MRI. However, while the sensitivity of DCE breast MRI reported in the literature is high (>90%), estimates of the specificity are much lower and vary widely. The use of DCE breast MRI is increasing and it is recommended as an adjunctive breast screening modality to mammography.2,22 DCE breast MRI techniques can detect some cancers missed by mammography23,24 and are especially beneficial for women with dense breast tissues, postoperative scars, or breast implants, or those at high risk for breast cancer.25,26 As an additional source of evidence for predicting lesion pathology and detecting multifocal disease, some studies suggest that DCE breast MRI is superior to nuclear medicine modalities;27–29 other studies, however, report that DCE MRI and nuclear medicine provide comparable accuracy.30 DCE breast MRI may be a powerful tool for predicting the likelihood of malignancy of very small lesions.31 It may also be useful for detecting and staging32,33 invasive lobular breast cancer,34 which can be particularly difficult to detect with conventional imaging. DCE MRI has also shown promise for predicting prognostic variables such as the presence of lymph node metastases,35–37 though not all studies have observed such a relationship.38 There is also evidence that it could be useful for assessing the lymph nodes directly by imaging the axilla.39 It has been suggested that DCE breast MRI may be useful in distinguishing between tumors that are or are not responding to neoadjuvant (presurgery) chemotherapy,40,41 though concerns have also been raised that changes in the contrast uptake behavior as a result of chemotherapy can lead to underestimation of the remaining tumor volume.42 Thus, there may be a role for DCE breast MRI in surgical planning,43 particularly since intraductal spread and presence of multifocal disease is better assessed by DCE MRI than by x-ray mammography.44,45 DCE MRI has also shown promise as an approach for monitoring residual tumors following surgical treatment.46 Of course, there are some drawbacks to DCE breast MRI relative to x-ray mammography. The likelihood of malignancy among breast lesions rated as “probably benign” by DCE MRI is substantially higher than for lesions rated as “probably benign” by mammography.47 While DCE breast MRI can reveal lesions that are not visible on x-ray mammography, it remains challenging to determine which of those lesions are likely to be malignant and thus warrant biopsy.48 In particular, lower accuracy has been reported for assessing microcalcifications.29 There are also lesions that are not detected at all on DCE breast MRI.49 The interpretation of breast MR images is a challenging task. Traditional manual interpretation of breast MRI is time consuming and tedious and can lead to oversight error due to the large size of 4D data sets (three spatial dimensions plus time). Manual interpretation is also subject to inter- and intraobserver variability, with some lesion characteristics (e.g., internal enhancement) showing considerably more observer variability than others (e.g., margin).50 There is a great need

744

C HAPTER 22

for computer-aided detection and diagnosis systems capable of increasing the efficiency, accuracy, and consistency of breast MRI interpretation.

22.4 C OMPUTER -A IDED D ETECTION

AND

D IAGNOSIS

The accuracy of image-based diagnosis is determined by both the image acquisition and the image interpretation processes. For example, missed lesions that are evident retrospectively are the cause of many of false-negative mammographic exams.8 Consequently, computer-aided detection (CADe) and diagnosis (CADx) systems for breast cancer have been the subject of considerable study. CADe systems are intended to aid radiologists in the process of detecting and localizing abnormalities on medical images and, as such, they help overcome perceptual errors such as oversight. CADx systems are intended to aid radiologists in the process of recommending an appropriate next action, typically biopsy or follow-up imaging, and, as such, they can help avoid missed cancers and obviate benign biopsies. It is important to note that CADe/CADx systems are designed to assist radiologists, not to perform autonomous diagnosis, although initial experiments typically focus on the independent performance of CADe/CADx systems. As well as the sensitivity or specificity improvements desired for any modality, CADe/CADx systems for DCE breast MRI are also needed to enable more efficient interpretation, thus reducing the high cost associated with breast MRI exams that has restricted their use in routine clinical practice. In fact, improving efficiency is the primary goal of current commercial systems for computer-aided DCE breast MRI interpretation. The majority of work to date in CADe/CADx has focused on x-ray mammography.51,52 CADe systems for x-ray mammography that have been approved by the U.S. Food and Drug Administration (FDA) include: Image Checker∗ ,53 MammoReader† ,54 and Second Look‡ .55 Other companies, such as VuComp, are in the investigational device phase at this time. There is evidence that current CADe systems increase the accuracy of mammography, at least for some types of breast lesions, although the extent of their impact is debated.56–61 CADx systems for mammography are not yet in routine clinical use, but have been the subject of numerous research studies. To the best of our knowledge, there are currently two commercial software systems for aiding with DCE breast MRI interpretation that are approved by the FDA: CADstream§62,63 and 3TP Software Option.¶64,65 While these systems represent an important step in CADe/CADx development, one should recognize that they are currently more limited than their more established counterparts in x-ray ∗ R2 Technologies, Sunnyvale, CA. † ISSI, Clearwater, FL. ‡ CADx Systems, Beavercreek, OH. Note that ISSI and Howtek (Hudson, NH) merged in 2002 to form iCAD and in 2003 CADx Systems also merged with iCAD (Nashua, NH). § Confirma™, Kirkland, WA. ¶ 3TP, New York, NY.

C OMPUTER -A IDED D IAGNOSIS

OF

B REAST C ANCER

ON

MR I MAGING

745

mammography. In particular, both systems currently on the market are aimed at improving efficiency only, although improving diagnostic accuracy may be a goal for CADe/CADx systems in the future. This important difference between the state of the art for DCE MRI and x-ray mammography is reflected in the kinds of approvals obtained from the FDA (please refer to Section 22.6.7). Thus, there is substantial room for improvement in DCE breast MRI CADe/CADx systems and more research is needed to improve algorithms at every stage of the processes.

22.5 D EVELOPING CAD E /CAD X

FOR

DCE B REAST MRI

In the following sections, the progress to date on developing CADe/ CADx systems for breast MRI is reviewed. The discussion is in terms of locating lesions on MRI and estimating their likelihood of malignancy, but very similar approaches would be used to develop a system that, for example, distinguished between lesions that were or were not responding to chemotherapy. In CADe for DCE breast MRI, the two key steps are image registration and lesion localization. The former is particularly challenging, while the latter is less complex in this modality since the region of enhancement is readily apparent. In CADx for DCE breast MRI, the same steps are needed as for CADe, with the addition of feature extraction and classification based on the extracted features. In almost all approaches, feature extraction requires that the lesion be segmented, not simply localized. Thus, in this presentation, segmentation is treated as an integral component of CADx. If a large number of features are extracted or if features are extracted with little a priori knowledge as to their likely value for the classification task, a feature-selection step can also be necessary. Again, in this presentation, the need for an explicit feature-selection stage is assumed. Finally, a few comments on issues in evaluating and validating CADe/CADx systems conclude the discussion. 22.5.1 I MAGE

REGISTRATION

It usually takes around 20–40 min to scan a sequence of breast MR images. During this relatively long acquisition process, respiratory and cardiac motion, as well as some degree of voluntary patient movement, are unavoidable. As a result of such displacements, the same coordinates in images at different times in the series may correspond to different physical locations in the subject. Thus, interpreting raw images can lead to errors in evaluating enhancement and morphology of an abnormality (Fig. 22.2). For this reason, one must compensate for the motion between the pre- and postcontrast images in order to achieve anatomical and functional correspondence. This process is referred to as registration. Although image registration is a traditional, well-explored topic in image analysis and computer vision and has been widely used in medical imaging applications,66,67 existing registration techniques cannot be simplistically applied to breast MR images. For example, some registration algorithms assume that the

746

C HAPTER 22

(a)

(b)

(c)

(d)

Figure 22.2 (a) Before motion. (b) After motion. (c) After subtracting (b) from (a) without registration. (d) Subtraction image following nonrigid registration of (a) and (b). Reproduced with permission from Ref. 68, Copyright IEEE.

image intensity is the same across the images to be registered, but this is not the case for DCE breast MR images. In fact, the key feature of DCE breast MRI is the enhancement of abnormal tissues after the administration of the contrast agent. A successful breast MRI registration technique must take into consideration that the breasts undergo nonrigid motion and that the image intensity changes over time. There are several algorithms that have been developed specifically for breast MRI registration. Hayton et al.69 use a modified Horn and Schunck optical flow model to minimize the fitting error of a pharmacokinetic model. The consistency of the pharmacokinetic model fitting is used as the similarity criterion in registration. Pharmacokinetic model failures can occur19 and such model failure may decrease the overall registration accuracy, as discussed by Rueckert et al.68

C OMPUTER -A IDED D IAGNOSIS

OF

B REAST C ANCER

ON

MR I MAGING

747

A feature-based 2D registration method is introduced in Lucht et al.70 In their approach, the correlation of edge features is used to maximize the similarity of precontrast and postcontrast images. However, the smoothness of the displacement vector field is not guaranteed. Since breast MRI is a 3D exam with high-spatial resolution, true 3D registration techniques would be preferred. Rueckert et al.68 introduce a nonrigid 3D registration method for breast MRI based on free-form deformations. The motions T (x, y, z) are modeled as a combination of global and local motions: T (x, y, z) = Tglobal (x, y, z) + Tlocal (x, y, z),

(22.1)

where the global component accounts for the overall motion of the human body, while the local component accounts for local deformations of the breast. The local motion model is a free-form deformation model based on B-splines to analyze the motion of the breast and the choice of cubic B-spline produces a smoothly varying displacement field. A general global transformation includes rotation, translation, scaling, and shearing, for a total of 12 degrees of freedom. A normalized mutual information criterion was chosen to maximize the similarity between the precontrast image and postcontrast image. The overall cost function is C = −Csimilarity + λCsmooth ,

(22.2)

in which the first term corresponds to the mutual information criterion and the second term corresponds to the smoothness cost associated with the deformation described by Eq. (22.1). An iterative method is used to find the optimal displacement vector. The constant λ is empirically optimized. Some results of this algorithm are reproduced in Fig. 22.2, which shows improvements in the image quality after registration. Rohlfing et al.71 extend the method of Rueckert et al.68 by using a local volume-preservation constraint to address the volume loss of contrast-enhancing structures, which is a limitation of the original algorithm. 22.5.2 L ESION

LOCALIZATION

After compensating for motion between precontrast and postcontrast images, subtracting the precontrast image from the first postcontrast image generates a subtraction image on which significant enhancement of the abnormality is obvious by visual inspection (Fig. 22.3). The region of interest (ROI) can be defined by placing a bounding box that completely contains the enhancement within the breast region and chest wall, as demonstrated by Hayton.72 Although, intuitively, one expects fewer false-positive detections than are typically found in the initial stages of CADe on mammography, false detections on DCE MRI could presumably result from blood flow in normal structures such as vessels. Additional study is needed to clarify this issue. A classifier could be developed to distinguish false detections

748

C HAPTER 22

(a)

(b)

(c)

Figure 22.3 Example of the enhancement region observed on a subtraction image formed by subtracting the precontrast image from the first postcontrast image. (a) Precontrast image. (b) First postcontrast image. (c) Subtraction image [(a) from (b)].

from true detections should that prove necessary. In that case, the stages of processing (lesion segmentation, etc.) as described for CADx in the following sections would be similarly used for false-positive reduction. 22.5.3 L ESION

SEGMENTATION

In order to compute morphological features and kinetic curves for use in making diagnosis and treatment decisions,70 the lesion must be accurately segmented from the ROI. Specifically, the goal of lesion segmentation is to partition the ROI into subregions in which the voxels share similar kinetic enhancement curves. Few previous publications are directly related to lesion segmentation on breast MRI. The two-level-thresholds method uses one threshold to segment the enhancement region from the background image and a second threshold to segment the malignant lesion from the enhancement region.73 The segmentation in Hayton’s work is also based on thresholding the enhancement rate72 of the first two images in the series. Since the signal intensity depends on the particular MRI instrumentation and contrast agent used in data acquisition, there is no general approach for selecting threshold values; thus, these methods require careful user interaction. Gihuijs et al.74 develop a seeds-based region-growing algorithm to segment the lesion from the ROI, using thresholds derived from the image histogram. However, the user must manually select the seeds. In addition, the same tissue may not behave uniformly in response to the contrast agent, which also decreases the accuracy of threshold-based methods. Since each of the kinetic enhancement curves is an n-dimensional vector, efforts have been made to cluster those curves such that the resulting clusters are the representatives of the partitioned subregions in the ROI. Petroudi et al.75 propose a Gaussian mixture model for dynamic breast MRI data, in which the mixture parameters are iteratively updated based on K-means initialization. However, K-means initialization is not stable and not recommended in general.

C OMPUTER -A IDED D IAGNOSIS

22.5.4 F EATURE

OF

B REAST C ANCER

ON

MR I MAGING

749

EXTRACTION

Once a lesion is detected in the ROI, characterization is necessary to estimate the pathological nature of the lesion, i.e., whether the lesion is benign or malignant. There is overlap in the appearance of kinetic enhancement curves of malignant and benign lesions (Fig. 22.4). Since both malignant and benign lesions can be enhanced, enhancement alone is not enough to determine the pathology of the lesion. In fact, one of the main criticisms of breast MRI has been its low specificity. It is generally recommended that both morphological features and kinetic enhancement patterns be employed in predicting the pathological nature of a lesion. The Breast Imaging Reporting and Data System (BI-RADS) is a lexicon for standardizing the description of morphological and kinetic features.77 Morphological features describe foci, masses, nonmasslike enhancements, and associated findings. Foci are small enhancements that can not be clearly described. Masses are described by shape and margin characteristics similar to those used for x-ray mammography with variations reflecting the differences in the modalities (e.g., a mass margin can not be “obscured” on MRI). Similarly, the internal enhancement characteristics of masses are described on MRI in analogy to mass density on mammography. On the other hand, the features describing nonmasslike enhancements and associated findings generally have less connection with the corresponding presentation on mammography. For example, the distribution of nonmasslike enhancement is described as focal area, linear enhancement, ductal enhancement, segmental enhancement, regional enhancement, or diffuse enhancement. The kinetic features in the BI-RADS lexicon are less developed at this time; the initial enhancement phase is simply reported as slow, medium, or fast and the delayed enhancement phase as persistent, plateau, or washout. The guidelines for wording the overall report acknowledge the fact that many people employ pharmacokinetic analyses, but no standardization of the approach is presented.

Figure 22.4 Ideal kinetic enhancement curve patterns of benign and malignant breast lesions on DCE MR. The horizontal axis represents time and the vertical axis represents image intensity. The curves labeled types Ia and type Ib are typical of benign lesions while those labeled type II and type III are typical of malignant lesions. (Image adapted from Ref. 76 with permission from the Radiological Society of North America.) Refer to the CD for the full color figure.

750

C HAPTER 22

BI-RADS does not specify an algorithm by which the lesion description should be mapped to a particular management strategy, but the lexicon provides a natural feature set that could be used for developing an automatic classifier for predicting appropriate next actions (e.g., perform biopsy because the lesion is likely malignant). Classification based on BI-RADS descriptors78–80 and automatic extraction of BI-RADS descriptors81,82 have been investigated for other modalities. Ideally, the extraction of lesion descriptions could be automated using imageprocessing methods. Several research groups have performed at least exploratory studies in this area (e.g., Refs. 31 and 83–86). Of particular interest is the increasing use of pharmacokinetic models to extract kinetic features.31,33,40,87 22.5.5 F EATURE

SELECTION

Feature selection is defined as a series of actions to choose a subset of features that are relevant to correct classification based on specified evaluation and selection criteria.88–90 There are several reasons why feature selection can be an important step in designing CADe/CADx systems. A classifier trained on a data set in which there are many more features than there are cases (patients) is less likely to perform adequately on new but supposedly similar cases; this phenomenon is referred to as “overtraining.”88–91 Thus, reducing the number of features can enable the system designer to build a more robust system and more accurately assess the performance of algorithms under consideration. The time it takes to train a classifier is typically dependent on the number of features, so decreasing the number of features can increase the efficiency of CADe/CADx system development. From a clinical perspective, it is desirable that the underpinnings of a CADe/CADx system be understandable by the human operator. A large number of redundant or irrelevant features are counter to this goal of transparency. Thus, it can be important to reduce the number of features from the set initially extracted. Feature selection has not been employed in all breast MRI CADe/CADx development studies. Presumably this is because many studies have used a small number of features that were designed to capture specific, intuitive properties of breast lesions on DCE MRI. However, some studies have used feature selection methods,31,36,83,84,92,93 particularly stepwise selection procedures31,36,83,84 such as in the context of regression models. It seems likely that as more image processing methods are applied to DCE MRI, the number of features to be considered will increase for two reasons. First, it is easier to collect numerous features when a computer calculates than when a human provides ratings. Second, there are general classes of image-based features that have shown moderate degrees of success in other applications that will presumably be explored further in DCE MRI in future studies. For example, Haralick’s texture features94 have been applied in only a few DCE MRI studies, such as that by Gibbs et al.95

C OMPUTER -A IDED D IAGNOSIS

22.5.6 L ESION

OF

B REAST C ANCER

ON

MR I MAGING

751

CLASSIFICATION

Some efforts have already been made to automatically predict lesion pathology based on morphological features, kinetic enhancement features, or both.73,83,84,86,93 Some such studies83,84 have demonstrated that a combination of morphological features and kinetic enhancement features provide better classification than either feature set alone. In most studies, features describing morphology or kinetic enhancement have been automatically extracted using image-processing methods. Several investigations of automatic lesion classification on DCE breast MRI as benign or malignant are summarized in Table 22.1 to provide an overview of this area of work. 22.5.7 CAD E /CAD X

EVALUATION AND VALIDATION

Rigorous validation of a CADe/CADx system is required in order to receive regulatory approval for clinical use, such as from the FDA. Current CADe systems for x-ray mammography are intended to impact the sensitivity and/or specificity and, as such, are regulated as “class 3” devices with classification of “analyzer, medical image.” Thus, premarket approval (PMA) was required for these systems. By comparison, the current software aids for DCE breast MRI are more limited in scope in that they emphasize improving the efficiency of the interpretation process. Consequently, the Confirma and 3TP systems are both essentially visualization aids and, as such, are regulated as “class 2” devices with classifications of “system, image processing, radiological” and “system, nuclear magnetic resonance imaging,” respectively. Thus, the less arduous 510(k) premarket notification was required for these systems. In summary, it is essential to improve both the efficiency and accuracy of DCE breast MRI interpretation, but systems to aid in these two tasks may be regulated differently. Validation of CADe/CADx systems should include evaluation of both the individual image analysis steps separately and of the overall CADe/CADx system. It is also necessary to investigate how performance at one step may influence later steps in the overall system.96 While in the early development stages it is common to evaluate CADe/CADx systems in a “stand-alone” mode, they ultimately must be tested with human observers.97 Previous work on developing CADe/CADx systems for other modalities and organ sites can provide a guide for researchers in breast MRI. Several recent articles present informative discussions of the issues in, and current approaches to, assessing CADe/CADx systems in medical imaging. The comprehensive review by Dodd et al.98 provides insight into subtle topics such as the complexities that can arise in defining “truth” for medical imaging analyses. Wagner et al.97 advocate the “multireader, multicase” paradigm as the “best practice” for assessing competing imaging modalities (including CADe/CADx). The review by Wirth99 is worthy of note because he summarizes the strategies that have been employed for evaluating

Segmentation SA

M

SA

M

M

M

M

Author Chen83

Gibbs31

Gilhujs84

Kinkel50

Lucht85

Szabo86

Vomweg93 M

A

A

M

A

A

Feature extraction A

Genetic algorithm

ARD

Feature selection Stepwise LDA Backward Elimination Stepwise LDA CART (embedded) N/A Morphological, kinetic Morphological, kinetic, clinical

Morphological, kinetic Morphological, kinetic Kinetic

Features Morphological, kinetic Kinetic

ANN

ANN

ANN

CART

LDA

LR

Classifier LDA

473/131

59/46

Number of cases (training/testing) 121 (leave-one out cross-validation) 43 (no crossvalidation) 27 (leave-one-out cross-validation) 57 (crossvalidation) 258/111

Sens. = 94% Spec. = 92%

Sens. = 91% Spec. = 83% Sens. = 84% Spec. = 81% Az = 0.77

Az = 0.96

Az = 0.92

Classifier performance Az = 0.86

Table 22.1 Summary of studies on developing classifiers to predict lesion pathology on breast MRI. A: automated; SA: semiautomated; M: manual; ANN: artificial neural network; LDA: linear discriminant analysis; LR: logistic regression; CART: classification and regression tree; ARD: automatic relevance determination.

752 C HAPTER 22

C OMPUTER -A IDED D IAGNOSIS

OF

B REAST C ANCER

ON

MR I MAGING

753

component algorithms such as registration and segmentation. It is also important to remember that many of the challenges in developing robust CADe/CADx systems are the same as for using statistical machine learning techniques in any context. Thus, general textbooks on classification methods can also be consulted.100,101 Here, three areas of particular concern based on the current state of the art in DCE breast MRI are highlighted. First, extensive databases in which the “ground truth” has been carefully verified are needed in order to compare CADe/CADx algorithms. To date, the majority of DCE breast MRI studies have enrolled few subjects (<100) and consequently have been exploratory in nature. One must be very cautious in applying algorithms for identifying subtle patterns to small data sets since they can identify spurious trends that do not hold up under additional scrutiny. Moreover, the performance of a system will show some variability even on databases of moderate size. Thus, it is not possible to fairly compare different algorithms unless they are evaluated on the same data set. Second, in order to meaningfully assess a predictive model, it is essential that the evaluation be performed using data that were not used to construct or train the model. This is typically achieved using some form of cross-validation in which a data set is partitioned into nonoverlapping subsets for the separate tasks of model training and testing. In some previous studies of simple classifiers for interpreting DCE breast MRI exams, it appears that the entire available data set was used for both model construction and evaluation. One needs to interpret the results of such studies cautiously since this approach can lead to optimistic assessments of the model performance. Third, consideration must be given to the choice of metric for evaluating the performance of a CADe/CADx system. For binary classification tasks (e.g., benign versus malignant), receiver operating characteristic (ROC) analysis is generally considered more clinically relevant than measures that depend on disease prevalence such as accuracy (percent correct) or mean-squared error. The area under the empirical curve (AUC) or under the curve fit based on a binormal model (Az ) is generally taken as a summary index. A particular operating point (sensitivity, specificity) associated with a particular threshold on the decision variable is also frequently reported, although comparisons can be unclear when both sensitivity and specificity vary. For an introduction to ROC analysis in medical imaging, please refer to Metz et al.102,103

22.6 F UTURE D IRECTIONS It is reasonable to suppose that organizational changes that support greater cooperation among research groups would lead to rapid developments in CADe/CADx for DCE breast MRI. Two key advancements of this nature are needed: standardization of the images to be analyzed and resources to support algorithm development. The adoption of a standard acquisition protocol for DCE breast MRI would facilitate the design of CADe/CADx systems. Of course, it would be desirable if

754

C HAPTER 22

CADe/CADx systems were robust to protocol changes, but that will likely be more difficult to achieve. Imaging standards need to address all aspects of the exam, not just acquisition parameters. For example, evidence is accumulating that enhancement patterns vary with a woman’s menstrual cycle87 and DCE MRI should be performed during the first half of the cycle. A publicly available database of breast MR images would enable researchers to perform preliminary, but direct, comparisons of alternate algorithms before embarking on larger, more costly studies. While some efforts have been made to establish public repositories of x-ray mammograms, notably the Digital Database for Screening Mammography ∗ and the Mammographic Image Analysis Society database,† similar efforts to establish a publicly available breast MRI database are unknown. There are important lessons to be learned from the experiences with mammography databases, or for that matter, any scientific databases that serve as public resources. First, creating a database is actually much easier than maintaining one. At a minimum, securing financial support for the latter can be difficult. Second, accurate “metadata” and explicit inclusion/exclusion criteria are critical to the community’s acceptance of a database. In this chapter, the discussion the stages of developing CADe/CADx systems as if they were independent, as has covered they are often developed independently. However, for example, the design choices for the registration and segmentation stages depend strongly upon each other. The dependencies among the different stages of CADe/CADx systems warrant greater attention in future research. The incorporation of high-level knowledge, such as the kinetics governing the accumulation of contrast agent, into registration and segmentation is a very promising direction for future work. In addition, kinetic information provides meaningful insight into the physiology of the tissue and can be incorporated as features for classification.104 As discussed earlier in this chapter, T2∗ -weighted images may provide supplemental information20,105 and their potential role in CADe/CADx systems warrants investigation. The overall performance of breast MRI CADe/CADx may be improved if information from T2∗ -weighted images were taken into account. Finally, novel image analysis techniques are needed to address the challenges unique to DCE MRI data, such as registration techniques that accommodate images with intensity changes and efficient methods for multidimensional image segmentation. Methods for fusing breast MRI data with x-ray mammography and other modalities would also be beneficial.

ACKNOWLEDGMENTS The authors would like to thank Chris Kite and Scott Swanson for technical assistance in the UT Biomedical Informatics Lab. ∗ http://marathon.csee.usf.edu/Mammography/Database.html. † http://www.wiau. man.ac.uk/services/MIAS/MIASweb.html.

C OMPUTER -A IDED D IAGNOSIS

OF

B REAST C ANCER

ON

MR I MAGING

755

R EFERENCES 1. Cancer Facts and Figures 2004. American Cancer Society, Atlanta (2004). 2. E.E. Penhoet, J.E. Joy, and D.B. Petitti (Eds.), Saving Women’s Lives: Strategies for Improving Breast Cancer Detection and Diagnosis, The National Academies Press, Washington, DC (2004). 3. J.T. Bushberg, J.A. Seibert, E.M. Leidholdt, Jr., and J.M. Boone, The Essential Physics of Medical Imaging, 2nd Ed., Lippincott Williams & Wilkins, Philadelphia (2002). 4. D.B. Kopans, Breast Imaging, Lippincott-Raven Publishers, Philadelphia (1998). 5. E.D. Pisano and M.J. Yaffe, “Digital mammography.” Radiology 234(2), p. 353 (2005). 6. K. Kerlikowske, P.A. Carney, B. Geller, M.T. Mandelson, S.H. Taplin, K. Malvin, V. Ernster, N. Urban, G. Cutter, R. Rosenberg, and R. BallardBarbash, “Performance of screening mammography among women with and without a first-degree relative with breast cancer.” Ann. Internal Med. 133(11), pp. 855–863 (2000). 7. T.M. Kolb, J. Lichy, and J.H. Newhouse, “Comparison of the performance of screening mammography, physical examination, and breast US and evaluation of factors that influence them: An analysis of 27,825 patient evaluations.” Radiology 225(1), pp. 165–175 (2002). 8. R.E. Bird, T.W. Wallace, and B.C. Yankaskas, “Analysis of cancers missed at screening mammography.” Radiology 184(3), pp. 613–617 (1992). 9. D.B. Kopans, “The positive predictive value of mammography.” Am. J. Roentgenology 158(3), pp. 521–526 (1992). 10. I. Reiser, R.M. Nishikawa, M.L. Giger, T. Wu, E. Rafferty, R.H. Moore, and D.B. Kopans, “Computerized detection of mass lesions in digital breast tomosynthesis images using two- and three-dimensional radial gradient index segmentation.” Technol. Cancer Res. Treatment 3(5), pp. 437–441 (2004). 11. D. Pavic, M. Koomen, C. Kuzmiak, and E.D. Pisano, “Ultrasound in the management of breast disease.” Current Women’s Health Reports 3(2), pp. 156– 164 (2003). 12. L. Weir, D. Worsley, and V. Bernstein, “The value of FDG positron emission tomography in the management of patients with breast cancer.” Breast J. 11(3), pp. 204–209 (2005). 13. W.B. Eubank and D.A. Mankoff, “Evolving role of positron emission tomography in breast cancer imaging.” Seminars Nucl. Med. 35(2), pp. 84–99 (2005). 14. J.P. Hornack, The Basics of MRI. http://www.cis.rit.edu/htbooks/mri/ (2005). 15. M.V. Knopp, F.L. Giesel, H. Marcos, H. von Tengg-Kobligk, and P. Choyke, “Dynamic contrast-enhanced magnetic resonance imaging in oncology.” TMRI 12(4), pp. 301–308 (2001).

756

C HAPTER 22

16. M.O. Leach, “Application of magnetic resonance imaging to angiogenesis in breast cancer.” BCR 3(1), pp. 22–27 (2001). 17. R. Matsubayashi, Y. Matsuo, G. Edakuni, T. Satoh, O. Tokunaga, and S. Kudo, “Breast masses with peripheral rim enhancement on dynamic contrast-enhanced MR images: correlation of MR findings with histologic features and expression of growth factors.” Radiology 217(3), pp. 841–848 (2000). 18. A.E. Merbach and E. Toth (Eds.), The Chemistry of Contrast Agents in Medical Magnetic Resonance Imaging, 1st Ed., Wiley, West Sussex (2001). 19. D.J. Collins and A.R. Padhani, “Dynamic magnetic resonance imaging of tumor perfusion.” IEEE Eng. Med. Biol. Mag. 23, pp. 65–83 (2004). 20. V.Y. Kuperman and M.T. Alley, “Differentiation between the effects of T1 and T2* shortening in contrast-enhanced MRI of the breast.” J. Magn. Reson. Imaging 9(2), pp. 172–176 (1999). 21. D.M. Ikeda, N.M. Hylton, K. Kinkel, M.G. Hochman, C.K. Kuhl, W.A. Kaiser, J.C. Weinreb, S.F. Smazal, H. Degani, P. Viehweg, J. Barclay, and M.D. Schnall, “Development, standardization, and testing of a lexicon for reporting contrast-enhanced breast magnetic resonance imaging studies.” J. Magn. Reson. Imaging 13(6), pp. 889–895 (2001). 22. S.J. Nass, I.C. Henderson, and J.C. Lashof (Eds.), Mammography and Beyond: Developing Technologies for the Early Detection of Breast Cancer, National Academies Press, Washington, DC (2001). 23. G.M. Kacl, P. Liu, J.F. Debatin, E. Garzoli, R.F. Caduff, and G.P. Krestin, “Detection of breast cancer with conventional mammography and contrastenhanced MR imaging.” Europ. Radiol. 8(2), pp. 194–200 (1998). 24. R.M. Warren and C. Hayes, “Localization of breast lesions shown only on MRI—a review for the UK Study of MRI screening for breast cancer, Advisory Group of MARIBS.” British J. Radiol. 73(866), pp. 123–132 (2000). 25. M.D. Schnall, “Application of magnetic resonance imaging to early detection of breast cancer.” Breast Cancer Res. 3(1), pp. 17–21 (2001). 26. C.K. Kuhl and H.H. Schild, “Dynamic image interpretation of MRI of the breast.” J. Magn. Reson. Imaging 12(6), pp. 965–974 (2000). 27. B. Bone, M.K. Wiberg, B.K. Szabo, A. Szakos, and R. Danielsson, “Comparison of 99mTc-sestamibi scintimammography and dynamic MR imaging as adjuncts to mammography in the diagnosis of breast cancer.” Acta Radiologica 44(1), pp. 28–34 (2003). 28. M. Heinisch, H.J. Gallowitsch, P. Mikosch, E. Kresnik, G. Kumnig, I. Gomez, P. Lind, H.W. Umschaden, J. Gasser, and E.P. Forsthuber, “Comparison of FDG-PET and dynamic contrast-enhanced MRI in the evaluation of suggestive breast lesions.” Breast 12(1), pp. 17–22 (2003). 29. G.L. Leinsinger, L. Friedl, R. Tiling, M.K. Scherr, D.T. Heiss, C. Kandziora, B. Camerer, H. Sommer, T. Pfluger, and K. Hahn, “Comparison of dynamic MR imaging of the breast and sestamibi scintimammography for evaluation of indeterminate mammographic lesions.” Europ. Radiol. 11(10), pp. 2050– 2057 (2001).

C OMPUTER -A IDED D IAGNOSIS

OF

B REAST C ANCER

ON

MR I MAGING

757

30. M. Imbriaco, S. Del Vecchio, A. Riccardi, L. Pace, F. Di Salle, F. Di Gennaro, M. Salvatore, and A. Sodano, “Scintimammography with 99mTc-MIBI versus dynamic MRI for non-invasive characterization of breast masses.” Europ. J. Nucl. Med. 28(1), pp. 56–63 (2001). 31. P. Gibbs, G.P. Liney, M. Lowry, P.J. Kneeshaw, and L.W. Turnbull, “Differentiation of benign and malignant sub-1 cm breast lesions using dynamic contrast enhanced MRI.” Breast 13(2), pp. 115–121 (2004). 32. P.J. Kneeshaw, L.W. Turnbull, A. Smith, and P.J. Drew, “Dynamic contrast enhanced magnetic resonance imaging aids the surgical management of invasive lobular breast cancer.” Europ. J. Surgical Oncology 29(1), pp. 32–37 (2003). 33. A. Qayyum, R.L. Birdwell, B.L. Daniel, K.W. Nowels, S.S. Jeffrey, T.A. Agoston, and R.J. Herfkens, “MR imaging features of infiltrating lobular carcinoma of the breast: histopathologic correlation.” Am. J. Roentgenology 178(5), pp. 1227–1232 (2002). 34. V. Cocquyt and S. Van Belle, “Lobular carcinoma in situ and invasive lobular cancer of the breast.” Current Opinion Obstetrics Gynecology 17(1), pp. 55– 60 (2005). 35. B.K. Szabo, P. Aspelin, M. Kristoffersen Wiberg, T. Tot, and B. Bone, “Invasive breast cancer: correlation of dynamic MR features with prognostic factors.” Europ. Radiol. 13(11), pp. 2425–2435 (2003). 36. N. Tuncbilek, H.M. Karakas, and O.O. Okten, “Dynamic magnetic resonance imaging in determining histopathological prognostic factors of invasive breast cancers.” Europ. J. Radiol. 53(2), pp. 199–205 (2005). 37. S. Mussurakis, D.L. Buckley, and A. Horsman, “Dynamic MR imaging of invasive breast cancer: correlation with tumour grade and other histological factors.” British J. Radiol. 70(833), pp. 446–451 (1997). 38. U. Fischer, L. Kopka, U. Brinck, M. Korabiowska, A. Schauer, and E. Grabbe, “Prognostic value of contrast-enhanced MR mammography in patients with breast cancer.” Europ. Radiol. 7(7), pp. 1002–1005 (1997). 39. A.D. Murray, R.T. Staff, T.W. Redpath, F.J. Gilbert, A.K. Ah-See, J.A. Brookes, I.D. Miller, and S. Payne, “Dynamic contrast enhanced MRI of the axilla in women with breast cancer: comparison with pathology of excised nodes.” British J. Radiol. 75(891), pp. 220–228 (2002). 40. M.D. Pickles, M. Lowry, D.J. Manton, P. Gibbs, and L.W. Turnbull, “Role of dynamic contrast enhanced MRI in monitoring early response of locally advanced breast cancer to neoadjuvant chemotherapy.” Breast Cancer Res. Treatment 91(1), pp. 1–10 (2005). 41. L. Martincich, F. Montemurro, G. De Rosa, V. Marra, R. Ponzone, S. Cirillo, M. Gatti, N. Biglia, I. Sarotto, P. Sismondi, D. Regge, and M. Aglietta, “Monitoring response to primary chemotherapy in breast cancer using dynamic contrast-enhanced magnetic resonance imaging.” Breast Cancer Res. Treatment 83(1), pp. 67–76 (2004).

758

C HAPTER 22

42. A. Rieber, H. Zeitler, H. Rosenthal, J. Gorich, R. Kreienberg, H.J. Brambs, and R. Tomczak, “MRI of breast cancer: influence of chemotherapy on sensitivity.” British J. Radiol. 70(833), pp. 452–458 (1997). 43. P.J. Drew, M.J. Kerin, T. Mahapatra, C. Malone, J.R. Monson, L.W. Turnbull, and J.N. Fox, “Evaluation of response to neoadjuvant chemoradiotherapy for locally advanced breast cancer with dynamic contrast-enhanced MRI of the breast.” Europ. J. Surgical Oncology 27(7), pp. 617–620 (2001). 44. S. Komatsu, C.J. Lee, Y. Hosokawa, D. Ichikawa, T. Hamashima, K. Shirono, H. Okabe, H. Kurioka, and T. Oka, “Comparison of intraductal spread on dynamic contrast-enhanced MRI with clinicopathologic features in breast cancer.” Japanese J. Clinical Oncology 34(9), pp. 515–518 (2004). 45. P.J. Drew, S. Chatterjee, L.W. Turnbull, J. Read, P.J. Carleton, J.N. Fox, J.R. Monson, and M.J. Kerin, “Dynamic contrast enhanced magnetic resonance imaging of the breast is superior to triple assessment for the preoperative detection of multifocal breast cancer.” Ann. Surgical Oncology 6(6), pp. 599–603 (1999). 46. D. Gianfelice, A. Khiat, M. Amara, A. Belblidia, and Y. Boulanger, “MR imaging-guided focused ultrasound surgery of breast cancer: correlation of dynamic contrast-enhanced MRI with histopathologic findings.” Breast Cancer Res. Treatment 82(2), pp. 93–101 (2003). 47. E.A. Sadowski and F. Kelcz, “Frequency of malignancy in lesions classified as probably benign after dynamic contrast-enhanced breast MRI examination.” J. Magn. Reson. Imaging 21(5), pp. 556–564 (2005). 48. K.C. Siegmann, M. Muller-Schimpfle, F. Schick, C.T. Remy, N. Fersis, P. Ruck, C. Gorriz, and C.D. Claussen, “MR imaging-detected breast lesions: histopathologic correlation of lesion characteristics and signal intensity data.” Am. J. Roentgenology 178(6), pp. 1403–1409 (2002). 49. A. Teifke, A. Hlawatsch, T. Beier, T. Werner Vomweg, S. Schadmand, M. Schmidt, H.A. Lehr, and M. Thelen, “Undetected malignancies of the breast: dynamic contrast-enhanced MR imaging at 1.0 T.” Radiology 224(3), pp. 881–888 (2002). 50. K. Kinkel, T.H. Helbich, L.J. Esserman, J. Barclay, E.H. Schwerin, E.A. Sickles, and N.M. Hylton, “Dynamic high-spatial-resolution MR imaging of suspicious breast lesions: diagnostic criteria and interobserver variability.” Am. J. Roentgenology 175(1), pp. 35–43 (2000). 51. M.L. Giger, “Computerized analysis of images in the detection and diagnosis of breast cancer.” Seminars Ultrasound 25(5), pp. 411–418 (2004). 52. M.P. Sampat, M.K. Markey, and A.C. Bovik, “Computer-aided detection and diagnosis in mammography.” Handbook of Image and Video Processing, A.C. Bovik (Ed.), pp. 1195–1217. Academic Press (2005). 53. U.S. Food and Drug Administration, Summary of Safety and Effectiveness Data: R2 Technologies. P970058 (1998). 54. U.S. Food and Drug Administration, Summary and Safety of Effectiveness Data: ISSI. P010038 (2002).

C OMPUTER -A IDED D IAGNOSIS

OF

B REAST C ANCER

ON

MR I MAGING

759

55. U.S. Food and Drug Administration, Summary of Safety and Effectiveness Data: CADx Medical Systems. P010034 (2002). 56. R.F. Brem, J. Baum, M. Lechner, S. Kaplan, S. Souders, L.G. Naul, and J. Hoffmeister, “Improvement in sensitivity of screening mammography with computer-aided detection: a multiinstitutional trial.” Am. J. Roentgenology 181(3), pp. 687–693 (2003). 57. D. Gur, J.H. Sumkin, L.A. Hardesty, and H.E. Rockette, “Computer-aided detection of breast cancer: has promise outstripped performance?” J. Natl. Cancer Inst. 96(9), pp. 717–718 (2004). 58. T.W. Freer and M.J. Ulissey, “Screening mammography with computer-aided detection: prospective study of 12,860 patients in a community breast center.” Radiology 220(3), pp. 781–786 (2001). 59. L.J. Warren Burhenne, S.A. Wood, C.J. D’Orsi, S.A. Feig, D.B. Kopans, K.F. O’Shaughnessy, E.A. Sickles, L. Tabar, C.J. Vyborny, and R.A. Castellino, “Potential contribution of computer-aided detection to the sensitivity of screening mammography.” Radiology 215(2), pp. 554–562 (2000). 60. D. Gur, J.H. Sumkin, H.E. Rockette, M. Ganott, C. Hakim, L. Hardesty, W.R. Poller, R. Shah, and L. Wallace, “Changes in breast cancer detection and mammography recall rates after the introduction of a computer-aided detection system?” (see comment). J. Natl. Cancer Inst. 96(3), pp. 185–190 (2004). 61. J.G. Elmore and P.A. Carney, “Computer-aided detection of breast cancer: has promise outstripped performance?” (comment). J. Natl. Cancer Inst. 96(3), pp. 162–163 (2004). 62. C. Wood, “Computer aided detection (CAD) for breast MRI.” Technol. Cancer Res. Treatment 4(1), pp. 49–53 (2005). 63. Summary of Safety and Effectiveness: Confirma (K013574), U.S. Food and Drug Administration, Editor (2003). 64. F. Kelcz, E. Furman-Haran, D. Grobgeld, and H. Degani, “Clinical testing of high-spatial-resolution parametric contrast-enhanced MR imaging of the breast.” Am. J. Roentgenology 179(6), pp. 1485–1492 (2002). 65. Summary of Safety and Effectiveness: 3TP (K031350), U.S. Food and Drug Administration, Editor (2003). 66. J.B. Maintz and M.A. Viergever, “A survey of medical image registration.” Med. Image Anal. 2(1), pp. 1–36 (1998). 67. J. Hajnal, D.J. Hawkes, and D. Hill (Eds.), Medical Image Registration, CRC Pess, Boca Raton (2001). 68. D. Rueckert, L.I. Sonoda, C. Hayes, D.L. Hill, M.O. Leach, and D.J. Hawkes, “Nonrigid registration using free-form deformations: application to breast MR images.” IEEE Trans. Med. Imaging 18(8), pp. 712–721 (1994). 69. P. Hayton, M. Brady, L. Tarassenko, and N. Moore, “Analysis of dynamic MR breast images using a model of contrast enhancement.” Med. Image Anal. 1(3), pp. 207–224 (1997).

760

C HAPTER 22

70. R. Lucht, M.V. Knopp, and G. Brix, “Elastic matching of dynamic MR mammographic images.” Magnetic Res. Med. 43(1), pp. 9–16 (2000). 71. T. Rohlfing, C.R. Maurer, Jr., D.A. Bluemke, and M.A. Jacobs, “Volumepreserving nonrigid registration of MR breast images using free-form deformation with an incompressibility constraint.” IEEE Trans. Med. Imaging 22(6), pp. 730–741 (2003). 72. P. Hayton, “Analysis of contrast-enhanced breast MRI.” Department of Engineering Science, University of Oxford (1998). 73. L. Arbach, A. Stolpen, and J.M. Reinhardt, “Classification of breast MRI lesions using a backpropagation neural network (BNN).” 2nd IEEE International Symposium on Biomedical Imaging: Macro to Nano, Piscataway, NJ (2004). 74. K.G.A. Gihuijs, M.L. Giger, and U. Bick, “A method for computerized assessment of tumor extent in contrast-enhanced MR images of the breast,” Computer-Aided Diagnosis in Medical Imaging, K. Doi et al. (Eds.), pp. 305– 310. Elsevier, Amsterdam (1999). 75. S. Petroudi, G. Ketsetzis, and M. Brady, “Multi-vector segmentation of breast MR images via hidden Markov random fields.” WSEAS Trans. Biology and Biomedicine (2004). 76. C.K. Kuhl, P. Mielcareck, S. Klaschik, C. Leutner, E. Wardelmann, J. Gieseke, and H.H. Schild, “Dynamic breast MR imaging: are signal intensity time course data useful for differential diagnosis of enhancing lesions?” Radiology 211(1), pp. 101–110 (1999). 77. American College of Radiology, ACR BI-RADS—Mammography, Ultrasound & Magnetic Resonance Imaging, Fourth Ed., American College of Radiology, Reston, VA (2003). 78. J.A. Baker, P.J. Kornguth, J.Y. Lo, and C.E. Floyd, Jr., “Artificial neural network: improving the quality of breast biopsy recommendations.” Radiology 198(1), pp. 131–135 (1996). 79. C.E. Floyd, Jr., J.Y. Lo and G.D. Tourassi, “Case-based reasoning computer algorithm that uses mammographic findings for breast biopsy decisions.” Am. J. Roentgenology 175(5), pp. 1347–1352 (2000). 80. G.D. Tourassi, M.K. Markey, J.Y. Lo, and C.E. Floyd, Jr., “A neural network approach to breast cancer diagnosis as a constraint satisfaction problem.” Med. Phys. 28(5), pp. 804–811 (2001). 81. M.P. Sampat, A.C. Bovik, and M.K. Markey, “Classification of mammographic lesions into BI-RADS shape categories using the beamlet transform.” Proceedings of SPIE 5741 (2005). 82. S. Paquerault, J. Yulei, R.M. Nishikawa, R.A. Schmidt, C.J. D’Orsi, C.J. Vyborny, and G.M. Newstead, “Automated selection of BI-RADS lesion descriptors for reporting calcifications in mammograms.”
,
, pp.
–
. 83. W. Chen, M.L. Giger, L. Lan, and U. Bick, “Computerized interpretation of breast MRI: investigation of enhancement-variance dynamics.” Med. Phys. 31(5), pp. 1076–1082 (2004).

C OMPUTER -A IDED D IAGNOSIS

OF

B REAST C ANCER

ON

MR I MAGING

761

84. K.G. Gilhuijs, M.L. Giger, and U. Bick, “Computerized analysis of breast lesions in three dimensions using dynamic magnetic-resonance imaging.” Med. Phys. 25(9), pp. 1647–1654 (1998). 85. R.E. Lucht, M.V. Knopp, and G. Brix, “Classification of signal-time curves from dynamic MR mammography by neural networks.” Magn. Reson. Imaging 19(1), pp. 51–57 (2001). 86. B.K. Szabo, P. Aspelin, and M.K. Wiberg, “Neural network approach to the segmentation and classification of dynamic magnetic resonance images of the breast: comparison with empiric and quantitative kinetic parameters.” Academic Radiology 11(12), pp. 1344–1354 (2004). 87. J.P. Delille, P.J. Slanetz, E.D. Yeh, D.B. Kopans, and L. Garrido, “Physiologic changes in breast magnetic resonance imaging during the menstrual cycle: perfusion imaging, signal enhancement, and influence of the T1 relaxation time of breast tissue.” Breast J. 11(4), pp. 236–241 (2005). 88. M. Dash and H. Liu, “Feature selection for classification.” Intell. Data Anal. 1(3) (1997). 89. M. Hall and G. Holmes, “Benchmarking attribute selection techniques for discrete data class data mining.” IEEE Trans. Knowledge Data Eng. (15) (2003). 90. I. Guyon and A. Elisseeff, “An introduction to variable and feature selection.” J. Mach. Learning Res. 3(7/8), pp. 1157–1182 (2003). 91. A.L. Blum and P. Langley, “Selection of relevant features and examples in machine learning.” Artif. Intell. 97(1-2), pp. 245–271 (1997). 92. B.K. Szabo, M.K. Wiberg, B. Bone, and P. Aspelin, “Application of artificial neural networks to the analysis of dynamic MR imaging features of the breast.” Europ. Radiol. 14(7), pp. 1217–1225 (2004). 93. T.W. Vomweg, M. Buscema, H.U. Kauczor, A. Teifke, M. Intraligi, S. Terzi, C.P. Heussel, T. Achenbach, O. Rieker, D. Mayer, and M. Thelen, “Improved artificial neural networks in prediction of malignancy of lesions in contrastenhanced MR-mammography.” Med. Phys. 30(9), pp. 2350–2359 (2003). 94. R.M. Haralick, “Statistical and structural approaches to texture.” Proceedings of IEEE 67(5), pp. 786–804 (1979). 95. P. Gibbs and L.W. Turnbull, “Textural analysis of contrast-enhanced MR images of the breast.” Magn. Res. Med. 50(1), pp. 92–98 (2003). 96. P. Jannin, J.M. Fitzpatrick, D.J. Hawkes, X. Pennec, R. Shahidi, and M.W. Vannier, “Validation of medical image processing in image-guided therapy.” IEEE Trans. Med. Imaging 21(12), pp. 1445–1449 (2002). 97. R.F. Wagner, S.V. Beiden, G. Campbell, C.E. Metz, and W.M. Sacks, “Assessment of medical imaging and computer-assist systems: lessons from recent experience.” Academic Radiology 9(11), pp. 1264–1277 (2002). 98. L.E. Dodd, R.F. Wagner, S.G. Armato, 3rd, M.F. McNitt-Gray, S. Beiden, H.P. Chan, D. Gur, G. McLennan, C.E. Metz, N. Petrick, B. Sahiner, J. Sayre, and G. Lung, “Image database consortium research, assessment methodologies and statistical issues for computer-aided diagnosis of lung nodules in

762

99. 100. 101. 102. 103. 104.

105.

C HAPTER 22

computed tomography: Contemporary research topics relevant to the lung image database consortium.” Academic Radiology 11(4), pp. 462–475 (2004). M.A. Wirth, “Performance evaluation of image processing algorithms in CADe.” Technol. Cancer Res. Treatment 4(2), pp. 159–172 (2005). T.M. Mitchell, Machine Learning. McGraw-Hill Series in Computer Science, C.L. Liu (Ed.), p. 414. WCB/McGraw-Hill, Boston (1997). R.O. Duda, P.E. Hart, and D.G. Stork, Pattern Classification, 2nd Ed., Wiley, New York (2000). C.E. Metz, “ROC methodology in radiologic imaging.” Investigative Radiol. 21(9), pp. 720–733 (1986). C.E. Metz, “Basic principles of ROC analysis.” Seminars Nucl. Medicine 8(4), pp. 283–298 (1978). P. Armitage, C. Behrenbruch, M. Brady, and N. Moore, “Extracting and visualizing physiological parameters using dynamic contrast-enhanced magnetic resonance imaging of the breast.” Med. Image Anal. 9(4), pp. 315–329 (2005). C.K. Kuhl, S. Klaschik, P. Mielcarek, J. Gieseke, E. Wardelmann, and H.H. Schild, “Do T2-weighted pulse sequences help with the differential diagnosis of enhancing lesions in dynamic breast MRI?” J. Magn. Reson. Imaging 9(2), pp. 187–196 (1999).