1887
banner image
No data available.
Please log in to see this content.
You have no subscription access to this content.
No metrics data to plot.
The attempt to load metrics for this article has failed.
The attempt to plot a graph for these metrics has failed.
Optimized approach to decision fusion of heterogeneous data for breast cancer diagnosis
Rent:
Rent this article for
USD
10.1118/1.2208934
    + View Affiliations - Hide Affiliations
    Affiliations:
    1 Department of Biomedical Engineering, Duke University, Durham, North Carolina 27705 and Duke Advanced Imaging Labs, Duke University, Durham, NC 27705
    2 Department of Biomedical Engineering, Duke University, Durham, North Carolina 27705 and Department of Electrical and Computer Engineering, Duke University, Durham, NC 27705
    3 Duke Advanced Imaging Labs, Department of Radiology, Duke University, Durham, North Carolina 27705
    4 Department of Biomedical Engineering, Duke University, Durham, North Carolina 27705 and Duke Advanced Imaging Labs, Duke University, Durham, NC 27705 and Medical Physics Graduate Program, Duke University, Durham, NC 27705
    a) Electronic mail: jonathan.jesneck@duke.edu
    Med. Phys. 33, 2945 (2006); http://dx.doi.org/10.1118/1.2208934
/content/aapm/journal/medphys/33/8/10.1118/1.2208934
http://aip.metastore.ingenta.com/content/aapm/journal/medphys/33/8/10.1118/1.2208934

Figures

Image of FIG. 1.
FIG. 1.

Feature group structure for calcification Data Set C (calcification lesions). The features of the calcification data set consisted of three main groups: Computer-extracted features, radiologist-extracted features, and patient history features. The computer-extracted features were morphological and shape features of the automatically detected and segmented microcalcification clusters within the digitized mammogram images. The radiologist-extracted features comprised both radiologist-interpreted findings and BI-RADS features. This data set consisted of pixel ROIs of all 1508 calcification lesions in the DDSM. This data set had many heterogenic characteristics, such as that it was collected at four different institutions, scanned on four digitizers with different noise characteristics, and included both human-extracted and computer-extracted features, such as shape and texture features.

Image of FIG. 2.
FIG. 2.

Feature group structure for mass Data set M (mass lesions). The features of the mass data set consisted of mammogram features, sonogram features, and patient history features. The mammogram features comprised both BI-RADS features and radiologist-interpreted findings. The sonogram features consisted of ultrasound BI-RADS features, Stavros features, and other ultrasound mass descriptors. All image features were radiologist-extracted features. The mass data set was heterogeneous in including both mammogram and sonogram views of the breast. Both mammogram and sonogram feature sets were as well as including patient history features.

Image of FIG. 3.
FIG. 3.

The role of likelihood-ratio thresholds for decision fusion. The first column shows plots of the log-likelihood-ratio versus feature value for each feature. The algorithm calculated the likelihood ratio and then thresholded it separately for each feature. The threshold determined the ROC operating point of the likelihood-ratio classifier of a particular feature. Next, the algorithm combined the binary decisions from the feature-level likelihood ratio classifiers using decision fusion theory to produce the likelihood ratio of the fused classifier.

Image of FIG. 4.
FIG. 4.

ROC curves for Data set C (calcification lesions). The classifiers’ ROC curves for 100-fold cross-validation are shown. Figure 4(a) shows the full ROC curves, while Figure 4(b) shows only the high-sensitivity region . For the calcification data set, the four classifiers yielded differing classification performance under 100-fold cross-validation. Both decision-fusion curves lay significantly above the LDA and ANN curves, both in terms of AUC and pAUC. As expected, the decision-fusion classifiers achieved the highest scores of all the classifiers for their target performance metrics; DF-A attained the greatest AUC, whereas DF-P attained the greatest pAUC. The DF-P curve surpassed the DF-A curve and dominated the other curves above the line . In order to gain high-sensitivity performance, DF-P sacrificed performance in the less clinically relevant range of .

Image of FIG. 5.
FIG. 5.

ROC curves for Data set M (mass lesions). For the mass data set, all classifiers had high levels of classification performance. The DF-A and DF-P achieved the highest AUC and pAUC, respectively. In terms of AUC, the DF-A outperformed both the ANN and LDA ( and 0.021, respectively). In (b), the DF-P curve had slightly more partial area than the other curves. Despite having statistically equivalent partial areas, the DF-P had a greater specificity than the LDA at high sensitivities .

Tables

Generic image for table
TABLE I.

Classifier performance on Data set C (calcification lesions). The table shows the AUC and pAUC values for the ROC curves of the four classifiers under 100-fold cross-validation. The performance values exhibited a wide range. The DF-A scored the best for AUC, while DF-P scored highest for pAUC, as expected. The decision-fusion curves soundly outperformed both the ANN and LDA in terms of pAUC.

Generic image for table
TABLE II.

values for AUC comparisons for Data set C (calcification lesions). The confusion matrix shows the values for the pairwise comparisons of the classifiers’ AUC values. All pairwise comparisons were statistically significant.

Generic image for table
TABLE III.

values for pAUC comparisons for Data set C (calcification lesions). The confusion matrix shows the values for the pairwise comparisons of the classifiers’ pAUC values. All pairwise comparisons were statistically significant.

Generic image for table
TABLE IV.

Classifier performance on Data set M (mass lesions). The table shows the AUC and pAUC values for the ROC curves of the four classifiers under 100-fold cross-validation. All four classifiers performed very similarly on this data set. The DF-A scored the best for AUC, whereas the DF-P scored highest for pAUC, although both were still within one standard deviation of each of the other classifiers’ performances.

Generic image for table
TABLE V.

values for AUC comparisons for Data set M (mass lesions). The confusion matrix shows the values for the pairwise comparisons of the classifiers’ AUC values. The DF-A outperformed the ANN and LDA. Among the DF-P, ANN, and LDA, there were no statistically significant pAUC differences.

Generic image for table
TABLE VI.

values for pAUC comparisons for Data set M (mass lesions). The confusion matrix shows the values for the pairwise comparisons of the classifiers’ pAUC values. None of the pAUC comparisons were statistically significant. Although pAUC scores were similar, the DF-P did have a higher specificity than the LDA at both 98% sensitivity versus and at 100% sensitivity versus ).

Loading

Article metrics loading...

/content/aapm/journal/medphys/33/8/10.1118/1.2208934
2006-07-26
2014-04-20
Loading

Full text loading...

This is a required field
Please enter a valid email address
752b84549af89a08dbdd7fdb8b9568b5 journal.articlezxybnytfddd
Scitation: Optimized approach to decision fusion of heterogeneous data for breast cancer diagnosis
http://aip.metastore.ingenta.com/content/aapm/journal/medphys/33/8/10.1118/1.2208934
10.1118/1.2208934
SEARCH_EXPAND_ITEM