Other
Breast Imaging
November 2000

Case-Based Reasoning Computer Algorithm that Uses Mammographic Findings for Breast Biopsy Decisions

Abstract

OBJECTIVE. We present case-based reasoning computer software developed from mammographic findings to provide support for the clinical decision to perform biopsy of the breast.
SUBJECTS AND METHODS. The case-based reasoning system is designed to support the decision to perform biopsy in those patients who have suspicious findings on diagnostic mammography. Currently, between 66% and 90% of biopsies are performed on benign lesions. Our system is designed to help decrease the number of benign biopsies without missing malignancies. Clinicians interpret the mammograms using a standard reporting lexicon. The case-based reasoning system compares these findings with a database of cases with known outcomes (from biopsy) and returns the fraction of similar cases that were malignant. This malignancy fraction is an intuitive response that the clinician can then consider when making the decision regarding biopsy.
RESULTS. The system was evaluated using a round-robin sampling scheme and performed with an area under the receiver operating characteristic curve of 0.83, comparable with the performance of a neural network model. If only the cases returning a malignancy fraction of greater than a threshold of 0.10 are sent to biopsy, no malignancies would be missed, and the number of benign biopsies would be decreased by 25%. At a threshold of 0.21, 98%, of the malignancies would be biopsied, and the number of benign biopsies would be decreased by 41%.
CONCLUSION. This preliminary investigation indicates that the case-based reasoning approach to computer-aided diagnosis has the potential to improve the accuracy of breast cancer diagnosis on mammography.

Introduction

We describe a new approach to computer-aided diagnosis of breast cancer using mammographic findings. Simplified case-based reasoning software has been developed to provide support for the clinical decision to perform breast biopsy. The system is designed to aid in the decision to perform a biopsy in patients who have suspicious mammographic findings. The decision to biopsy can be viewed as a two-stage process. First, the mammographer views the mammogram and determines the presence or absence of image features such as calcifications and masses. Second, the presence and description of these features and the patient's medical history are merged to form a diagnosis. The case-based reasoning system is an aid to the second step and is motivated by the large fraction of biopsies that are benign.
Mammography is a sensitive procedure for detecting breast cancer, but the positive predictive value is low. Only 10-34% of women who undergo biopsy for mammographically suspicious impalpable lesions are actually found to have malignancy [1]. Between 0.5% and 2.0% of all mammographic examinations result in biopsy; several hundreds of thousands of biopsies are performed on benign lesions each year. The women undergoing biopsy for a benign finding are unnecessarily subjected to the discomfort, expense, potential complications, change in cosmetic appearance, and anxiety that can accompany breast biopsy [1,2,3,4]. In addition, the financial burden of these procedures ($3000-$5000 per biopsy) is significant in the present political and economic effort to reduce expenditures. The proposed system may significantly improve this performance through a simplified case-based reasoning approach that uses a large database of cases with known outcomes. In clinical practice, this system can be easily integrated into the mammographers' work flow through a computerized reporting system. The clinician interprets a mammogram and records the findings into a computer using a standard reporting lexicon (Breast Imaging Reporting And Data System [BI-RADS]) [5]. The database is searched for similar cases, and the fraction of similar cases that were malignant is returned. This fraction is referred to as the “malignancy fraction” and is an intuitive response that the woman's health care team can then include in the medical decision for biopsy.

Subjects and Methods

The case-based reasoning system functions as a database query of mammographic cases with known biopsy outcomes. When a new case (called a “test case”) is considered for biopsy, the case-based reasoning system will match the test case to the entire database. The case-based reasoning system returns the ratio of the number of malignant to the total number of cases that match. This malignancy fraction is intuitively satisfying as an indicator for diagnosis of malignancy.
The critical components of a case-based reasoning system are as follows: a translatable quantitative encoding for cases, a database of cases, and a rule to define when the test case matches a case in the database.
The input to the system is a subset of the BI-RADS [5] mammographic findings and items from the patients' medical histories. The output is formed from the known outcome of biopsy. The database consists of a set of radiologist's findings, a medical history, and a set of biopsy results for cases that were judged suspicious for breast cancer and for which excisional biopsies were obtained as the gold standard. For the work reported here, an existing set of 500 cases was used. The heart of the system is the set of rules defining a match between the test case and the known cases in the database. The trivial match criterion would be that all findings match exactly. Although easy to justify, this matching rule is impractical. Considering 10 of the categoric BI-RADS findings, there are more than 1010 combinations of features; exact matches are rare. Cases are matched if they are “close” in some sense. For this work, a mathematic form of this “closeness” was developed using an ad hoc deterministic matching rule described in a subsequent section.

Cases

The cases for this project were described for a previous investigation to develop an artificial neural network for the decision to biopsy [6].
Of the women undergoing needle localization for impalpable breast lesions between January 1991 and December 1995, 500 lesions were randomly selected for open excisional biopsy and pathologic diagnosis. These lesions include 206 that were retrospectively interpreted in a previous study [7] and 294 new cases that were prospectively acquired.
Each set of mammograms was acquired using film-screen technique on dedicated mammography equipment. No case was included in the study if either of the reviewing radiologists had prior knowledge of the biopsy results or if the suspicious area was not definitely identified. Of the 500 lesions evaluated there were 232 masses alone, 192 microcalcifications alone, and 29 combinations of masses and associated microcalcifications. The remaining 47 lesions included various combinations of architectural distortion, regions of asymmetric breast density, areas of focal asymmetric density, and areas of asymmetric breast tissue. Patients were 24-86 years old (average age, 55 years). At biopsy, 326 (65%) of the lesions were found to be benign, and 174 (35%) were malignant. This positive predictive value of 35% is greater than those reported in prior studies [1, 4, 8, 9] but is consistent with our previous data.
All mammograms were interpreted by radiologists whose primary clinical responsibilities are the interpretation of mammograms and the evaluation of breast lesions and who routinely report case findings using the BI-RADS [5] descriptors. Both radiologists were asked to describe each lesion using the BI-RADS lexicon by completing a checklist that included all possible BI-RADS descriptors. Both radiologists were permitted to select only a single descriptor from each category. The findings were recorded during the routine patient workup before biopsy results were known. The reviewing mammographer was provided with the patient's history and any prior films.
The cases were randomly numbered with no identifying marks that could be traced to the original patients to ensure that patient confidentiality was maintained.

Input Findings

The input features were selected from 10 of the features from the BI-RADS [5] lexicon and one finding from the medical history. The 10 features initially considered from the BI-RADS lexicon were chosen on the basis of our previous work with these data [6] and included mass size, mass margin, mass density, mass shape, calcification description, calcification number, calcification distribution, and special cases or associated findings. The patient's age was included among the history findings. We found that performance strongly depended on which features were included in the matching criteria. No sophisticated feature-selection algorithm was used. To reduce the initial number of features, a forward stepwise linear discriminant analysis was performed with these 11 potential input features, and six were found to contribute at a significance level of 0.05 or greater. These selected features were age, mass margin, mass density, calcification description, calcification distribution, and associated findings (including the architectural distortion descriptor). The case-based reasoning system can be considered a very restricted linear model; therefore, feature exclusion using linear discriminant analysis should retain any features useful for case-based reasoning. In addition, the mammographer rated the likelihood of malignancy on a 5-point scale for each case. This rating was independent of the BI-RADS assessment.

Case-Based Reasoning Algorithm

Given a test case, a database, and a matching rule, the case-based reasoning algorithm is straightforward. From all cases in the database, those that match the test case are selected, and then the malignancy fraction is computed as the number of selected cases that have malignant outcomes divided by the total number of selected cases.

Matching Rules

The most important component of this system is the rule to decide whether a new case is similar to a case in the database. The simplest matching rule is to accept all cases. The next logical matching rule is to require that the type of lesion be matched: mass or calcification. Next, we require a match for the primary finding: mass margin for masses or calcification description for calcifications. Then we consider all findings. By trying combinations of findings one can obtain an optimized matching rule. We define a distance function as the total number of features that do not match between the test case and the example from the database. This rule introduces a parameter: the distance cutoff. If an example in the database has a distance from the test case that is less than the distance cutoff, then it matches.

Case Indexing

In formal case-based reasoning, case indexing refers to techniques that allow rapid identification and retrieval of similar cases. For this problem, the small number of features per case, combined with the inherent categoric structure of the features, allows efficient implementation while considering all features of every case. Only the patient age and the mass size are not categoric. A sliding window was used to match patient age and mass size. If the difference in feature value between two cases is less than the window width, then the features match.

Sampling for Testing

Currently, much has been written in the statistics literature concerning sampling from a finite number of examples to evaluate the performance of modeling systems. Following the work of Tourassi et al. [10] with neural network system training and testing, we adopted a round-robin technique. In this technique, a testing set is formed by removing only one of the examples. The system is built from the remaining examples and then tested on the one that was removed. The testing example is replaced in the set and another is removed. The system is again built and tested. This procedure is repeated until all examples have been used as testing cases. Performance of the case-based reasoning system is evaluated by setting a threshold on the malignancy fraction for the set of all tested examples. By adjusting this threshold on the malignancy fraction over the range from zero to one and computing the sensitivity and specificity of the system at each threshold, the receiver operating characteristic (ROC) curve is generated for analysis.

Analysis

The performance of the system was quantified using the area under the ROC curve, the area under the partial ROC curve, and specificity at a fixed sensitivity. It is general practice to use a fitting algorithm to estimate the area under the ROC curve. The most recognized of these programs are those from Charles Metz [11]. Unfortunately, the program (ROCKIT; Charles Metz, University of Chicago, Chicago, IL) produced a fit that did not well represent the actual data in the region of greatest interest: the high-sensitivity region. For this reason, the areas under the curves reported here were computed directly from the data using Newton's method of integration. As an unfortunate consequence of using the data directly, we are currently unable to report the statistical significance of differences between two ROC curves. Although the area under the ROC curve is a customary measure of performance, the weights assigned to sensitivity and specificity are equal. For breast cancer prediction, the cost of missing a malignancy is greater than the cost of performing a benign biopsy; therefore, a more appropriate measure is the performance at high sensitivity. For this reason, we report three other measures: the area under the partial ROC curve reported for a sensitivity of 0.9 or greater, the specificity at a sensitivity of 100% (representing the fraction of benign biopsies that could be saved while missing no malignancies), and the specificity at a sensitivity of 98% (representing the fraction of benign biopsies that could be saved while missing 2% of the malignancies). Note that perfect performance has an area under the full ROC curve of 1.0 and that chance behavior has an area of 0.5. For the partial ROC curve, perfect performance has an area of 0.1, and chance behavior has an area of 0.005.

Results

The performance of the system using different matching rules is shown in Table 1 and is described in a subsequent section. The simplest matching rule is comparing a new case to all cases (shown as rule “All cases” in Table 1), which gives the obvious performance of returning just the positive predictive value of the database (0.35 for our data). There is no decision variable, and thus no area under the ROC curve is computed. Requiring that the type of lesion be matched (mass or calcification) and setting a threshold on the malignancy fraction (shown as rule “Lesion type” in Table 1) gave an area of 0.37. Requiring a match for only the mass margin for masses or calcification description for calcifications (shown as rule “1 Finding” in Table 1) gave an area of 0.40. All these results are worse than chance behavior.
TABLE 1 Performance of Case-Based Reasoning with Different Matching Rules Compared with Artificial Neural Network (ANN)
Matching RuleDistanceReceiver Operating Characteristic (ROC) AreaPartial ROC AreaSpecificity at 100% SensitivitySpecificity at 98% Sensitivity
All casesNANANA00
Lesion type00.37<0.005<0.01<0.01
1 Finding00.40<0.005<0.01<0.01
6 Findings00.70<0.005<0.01<0.01
6 Findings10.790.02<0.01<0.01
3 Findings10.830.0450.250.41
ANN
NA
0.86
0.048
0.07
0.26
Note.—NA = not applicable.
As a next refinement, we considered matching exactly those six findings found to be significant by linear discriminant analysis (shown as rule “6 Findings,” “Distance = 0” in Table 1). This gave an area of 0.7 but with poor specificity (<1%) at a high sensitivity. Next, one mismatched finding was allowed (shown as rule “6 Findings,” “Distance = 1” in Table 1) resulting in an area of 0.79. Although this area under the ROC curve is encouraging, the performance at a high sensitivity was disappointing, with an area under the partial ROC curve of only 0.02 and a specificity of less than 1% at both 100% and 98% sensitivities.
Various combinations of the six features were then investigated. Although not every possible combination of features was investigated, a human-directed search resulted in optimal performance using only three features: mass margin, calcification description, and age with a matching window of 3 years difference. A case was allowed to match if two or three of the three features matched. Shown as rule “3 Findings,” “Distance = 1” in Table 1, this system performed with an area of 0.83, a partial area of 0.045, a specificity of 25% at 100% sensitivity, and a specificity of 41% at 98% sensitivity. This is comparable with the area of 0.86 that has been reported for artificial neural networks and for mammographers on a subset of this database [12].
The performance of the case-based reasoning system can be compared with that of an artificial neural network model developed from these same data [6] (shown as “ANN” in Table 1). For a round-robin evaluation, the artificial neural network had an area under the full ROC curve of 0.86 compared with 0.83 for the case-based reasoning system. The partial area for the artificial neural network was 0.048 compared with 0.046 for the case-based reasoning system. The specificity at 100% sensitivity was 0.30 compared with 0.25 for the case-based reasoning system. The specificity at 98% sensitivity was 0.42 compared with 0.41 for the case-based reasoning system. These results are similar, with no statistically significant differences.
All the following results are for the case-based reasoning system with three inputs and with a distance cutoff of one. The histogram for this best performance shows the separation of the malignant and benign cases when malignancy fraction is used as a decision variable (Fig. 1). The cases are binned by malignancy fraction from the round-robin test. The bin labels correspond to the left edge of each bin. Thus, the bin labeled 0 includes cases with malignancy fractions of 0 up to but not including 0.1 and has 80 benign and no malignant cases, and the bin labeled 0.1 has 44 benign cases and three malignant cases.
Fig. 1. Histogram shows benign and malignant cases binned by malignancy fraction output of case-based reasoning system. Note that although benign and malignant cases overlap considerably, 81 cases with malignancy fraction of less than 0.1 are all benign. These patients could be spared surgery without missing any malignancies. Dark shading indicates malignant; light shading indicates benign.
Examination of the histogram reveals that the distribution of benign cases is bimodal and that of the malignant cases is not and thus violates the assumptions of the maximum likelihood fitting algorithms typically used to analyze diagnostic systems. The expanded histogram (Fig. 2) shows that if a threshold were set at a value of 0.1 for the output of the case-based reasoning system, then all malignancies could be identified to the right of the threshold (justifiably biopsied), and 81 of the 326 benign cases could be identified to the left (avoided unnecessary biopsy).
Fig. 2. Histogram of benign and malignant cases binned by malignancy fraction expanded around high-sensitivity region binned by malignancy fraction. Note that if threshold were set at 0.15, additional 40 benign cases would not have been biopsied, and one malignancy would have been missed. Dark shading indicates malignant; light shading indicates benign.
An encouraging aspect of the performance of the case-based reasoning system is shown by the performance at high sensitivity. As can be seen in Table 2, 100% sensitivity can be maintained with a specificity of 25%. This represents a positive predictive value improvement from 35% to 42%. Of the 326 benign biopsies, 81 patients would have been spared surgery. With a sensitivity of 98%, specificity is 41%. This represents a positive predictive value improvement from 35% to 46%. Of the 326 benign biopsies, 134 patients would have been spared surgery and 10 of the 174 malignancies would have been missed.
TABLE 2 Performance of Best Case-Based Reasoning System as a Function of Threshold
No. of Malignancies MissedNo. of Benign Biopsies AvoidedSensitivity (%)Specificity (%)Positive Predictive Value (%)Threshold
00100035Nonea
08110025420.10b
10
134
98
41
46
0.21c
a
No use of system.
b
Corresponds to 100% sensitivity.
c
Corresponds to 98% sensitivity.
As Table 2 shows, to use the system at 100% sensitivity, any case returning a malignancy fraction of greater than 0.1 would be biopsied. To use the system at 98% sensitivity, any case returning a malignancy fraction of greater than 0.21 would be biopsied. Another way to interpret the output is to enter the case findings and examine the malignancy fraction. If the malignancy fraction for similar cases is greater than or equal to 0.1 (indicating that ≥10% of the cases that were similar to the new case were malignant), then a biopsy should be performed and the sensitivity of this decision should be 100% with a specificity of 25%.

Examples

In one case (Fig. 3) a 73-year-old woman's mammogram shows a region with a cluster of more than 10 fine-branching microcalcifications. The mammographer indicated that this case was very likely malignant. Sixty cases were found to match, and the malignancy ratio was 0.53. With a threshold of 0.10, this ratio of 0.53 would indicate biopsy. The histologic diagnosis for this case was malignant.
Fig. 3. Mammogram of 73-year-old woman with cluster of microcalcifications (>10) described as fine-branching. Cluster is at tip of white arrowhead. Both mammographer and case-based reasoning would correctly recommend biopsy of this malignant lesion.
In another case (Fig. 4), a 43-year-old woman's mammogram shows an isodense lobulated 18-mm mass with a well-circumscribed margin. The mammographer indicated that this case was very likely benign. The findings from this case matched 121 cases in the database, resulting in a malignancy ratio of 0.07. With a threshold of 0.10, no biopsy would be indicated. The histologic diagnosis for this case was benign.
Fig. 4. Mammogram of 43-year-old woman with isodense lobulated 18-mm mass with well-circumscribed margin. Mass is at tip of white arrowhead. Both mammographer and case-based reasoning would correctly recommend no biopsy of this benign lesion.
In a third case (Fig. 5), a 45-year-old woman's mammogram shows an isodense irregular 25-mm mass with an obscured margin. The mammographer indicated that this case was very likely malignant. Of the 88 cases that matched this third case, nine were malignant, producing a ratio of 0.10. With the threshold lowered to 0.15, the system would not recommend biopsy, and this malignancy would have been missed if the decision to biopsy were made solely by the computer. The histologic diagnosis for this case was malignant. This example represents the malignant case with the lowest malignancy ratio (0.10) returned by the case-based reasoning system. This is the malignant case that would not be recommended for biopsy if the threshold were raised to 0.15 to save 40 more benign biopsies. The mammographer indicated that this case was very likely malignant. In such a situation, the opinion of the mammographer would be accepted, and the patient would be correctly referred for biopsy.
Fig. 5. Mammogram of 45-year-old woman with isodense irregular 25-mm mass with obscured margin. Mass is at tip of white arrowhead. Mammographer correctly recommended biopsy of this malignant lesion. Case-based reasoning would miss this malignancy if threshold were increased to avoid 40 more biopsies.

Computational Time

With a database of 500 cases, a new case can be compared with the entire database in less than 0.04 sec running in a nonoptimized database language on a Pentium III 600-MHz personal computer (Intel, Santa Clara, CA).

Discussion

Other researchers, including those who proposed rule-based systems [13], artificial neural network approaches [6, 14, 15], and, recently, Bayesian networks [16], have worked on this problem. Rule-based systems are attractive because the reasoning behind each decision can be readily explained. The difficulty lies with the formulation of the rules and with the multiple exceptions that are usually required. In addition, no rule-based system has reported performance comparable with the case-based reasoning results presented here.
Although the artificial neural network techniques show excellent performance, interpretation of an individual prediction is challenging because of the nonlinear multidimensional representation of the decision space. Comparison of case-based reasoning performance with that of an artificial neural network model that was optimized with the same findings and cases shows that although the ROC area of the artificial neural network is larger, the specificity of the case-based reasoning system was better at high sensitivity when such a decision aid is likely to be used. The Bayesian network approach is theoretically attractive, although the underlying assumptions may be difficult to justify given the finite size of available databases. This case-based reasoning approach is easily implemented directly in the relational databases on which current mammography reporting systems are built.
A major attraction of this technique is its simplicity and intuitive clarity. The case-based reasoning system estimates the answer to the question, “Of all cases that are similar to this one, how many were malignant at biopsy?” The mechanism is easy to understand: Find all of the previous cases that are similar, and then report the fraction of those cases that were malignant.
One disadvantage of the case-based reasoning technique is the possibility that a new case will be presented that has no match in the database. Although this did not occur in the current database with the coarse matching criteria described previously, it is more likely to occur as more restrictive matching criteria are applied. This can be addressed by an adaptive matching criteria that will broaden the criteria for a case match, if too few matches are found, using a more strict criteria. Expansion of the database will also decrease the probability of such an occurrence.
Another potential difficulty is that two mammographers sometimes will report the same lesion using different descriptors. This difficulty may be overcome by matching rules that allow two cases to be matched if they are similar but do not require that they are identical. An example of this is the criteria described in this section for the age feature where a match was formed if two ages were within 3 years. Two interpretations of the same mammogram might not be identical, but they are likely to be similar.
The proposed system has the potential to add much practical value to a mammography reporting system. One use of the system is to serve as reference material for the mammographer who is considering a case for biopsy. A mammographer in a busy medical center may see 600 biopsied cases in a year. Given a 35-year career, this hypothetic mammographer may see fewer than 20,000 cases referred for biopsy. A more typical mammographer would see fewer than 200 cases per year with a career total of 7000. It is reasonable to expect that the system described here could easily contain more cases than a typical or even a busy mammographer could see during a lifetime of work. Evaluating a new case against such a database of 25,000 cases could be performed in less than 2 sec using a 600-MHz personal computer. The prediction system can be integrated in a seamless manner because most mammography reporting systems are database applications. Such a system could report the malignancy fraction to the mammographer less than 4 sec after the findings for the case were entered into the reporting system.
For the threshold of 0.1, we reported 81 benign biopsies that could have been avoided. These included 60 masses with well-circumscribed margins, one mass with a microlobulated margin, 18 masses with obscured margins, one mass with an ill-defined margin and with associated calcifications described as coarse, and one mass with a well-circumscribed margin and with associated calcifications described as indistinct.
The study described in this article considered only cases in which biopsies were performed. The actual sensitivity of the mammographers in this database was thus 100%, and their specificity was 0% because every case was biopsied. In a future study, we will include cases in which patients were considered for biopsy but were followed up instead. Some fraction of these patients will have developed cancer that would have been diagnosed had a biopsy been performed earlier. This future evaluation will examine whether the case-based reasoning system would have found any of these malignancies and thus improve the sensitivity of mammography.

Footnotes

The opinions and assertions contained herein are the private views of the authors and are not to be construed as official or as reflecting the views of the Department of the Army or the National Institutes of Health.
Supported by grant R21-CA81309 awarded by the National Cancer Institute, National Institutes of Health, and by grant DAMD17-99-9174 awarded by the United States Army Medical Research Acquisition Activity.
Address correspondence to C. E. Floyd, Jr.

References

1.
Kopans DB. The positive predictive value of mammography. AJR 1992; 158:521-526
2.
Dixon JM, John TG. Morbidity after breast biopsy for benign disease in a screened population. Lancet 1992; 1:128
3.
Helvie MA, Ikeda DM, Adler DD. Localization and needle aspiration of breast lesions: complications in 370 cases. AJR 1991; 157:711-714
4.
Schwartz GF, Carter DL, Conant EF, Gannon FH, Finkel GC, Feig SA. Mammographically detected breast cancer: nonpalpable is not a synonym for inconsequential. Cancer 1994; 73:1660-1665
5.
Lo JY, Baker JA, Kornguth PJ, Floyd CE Jr. Effect of patient history findings on predicting breast cancer from mammograms using artificial neural networks. Acad Radiol 1999; 6:10-15
6.
American College of Radiology. Breast imaging reporting and data system, 3rd ed. Reston, VA: American College of Radiology, 1998
7.
Baker JA, Kornguth PJ, Lo JY, Floyd CE Jr. Artificial neural network: improving the quality of breast biopsy recommendations. Radiology 1995; 198:131-135
8.
Ciatto S, Cataliotti L, Distante V. Nonpalpable lesions detected with mammography: review of 512 consecutive cases. Radiology 1987; 165:99-102
9.
Hall FM, Storella JM, Silverstone DZ, Wyshak G. Nonpalpable breast lesions: recommendations for biopsy based on suspicion of carcinoma at mammography. Radiology 1988; 167:353-358
10.
Tourassi GD, Floyd CE Jr, Sostman HD, Coleman RE. Artificial neural network for diagnosis of acute pulmonary embolism: effect of case and observer selection. Radiology 1995; 194:889-893
11.
Metz CE. ROC methodology in radiologic imaging. Invest Radiol 1986; 21:720-733
12.
Baker JA, Kornguth PJ, Lo JY, Williford ME, Floyd CE Jr. Breast cancer: prediction with artificial neural network based on BI-RADS standardized lexicon. Radiology 1995; 196:817-822
13.
Getty DJ, Pickett RM, D'Orsi CJ, Swets JA. Enhanced interpretation of diagnostic images. Invest Radiol 1988; 23:240-252
14.
Wu Y, Giger ML, Doi K, Vyborny CJ, Schmidt RA, Metz CE. Artificial neural networks in mammography: application to decision making in the diagnosis of breast cancer. Radiology 1993; 187:81-87
15.
Floyd CE Jr, Lo JY, Yun AJ, Sullivan DC, Kornguth PJ. Prediction of breast cancer malignancy using an artificial neural network. Cancer 1994; 74:2944-2948
16.
Kahn CE Jr, Roberts LM, Shaffer KA, Haddawy P. Construction of a Bayesian network for mammographic diagnosis of breast cancer. Comput Biol Med 1997; 27:19-29

Information & Authors

Information

Published In

American Journal of Roentgenology
Pages: 1347 - 1352
PubMed: 11044039

History

Submitted: January 22, 1999
Accepted: May 3, 2000

Authors

Affiliations

Carey E. Floyd, Jr.
Department of Radiology, Duke University Medical Center, Box 2623, Durham, NC 27710.
Department of Biomedical Engineering, Duke University Medical Center, Durham, NC 27710.
Joseph Y. Lo
Department of Radiology, Duke University Medical Center, Box 2623, Durham, NC 27710.
Department of Biomedical Engineering, Duke University Medical Center, Durham, NC 27710.
Georgia D. Tourassi
Department of Radiology, Duke University Medical Center, Box 2623, Durham, NC 27710.

Metrics & Citations

Metrics

Citations

Export Citations

To download the citation to this article, select your reference manager software.

Articles citing this article

View Options

View options

PDF

View PDF

PDF Download

Download PDF

Media

Figures

Other

Tables

Share

Share

Copy the content Link

Share on social media