Bei Liu PhD, Presenter: Nothing to Disclose

Charles Edgar Metz PhD, Abstract Co-Author: Nothing to Disclose

Yulei Jiang PhD, Abstract Co-Author: Nothing to Disclose

To compare the performance of three methods of combining information derived from two images of the same patient (e.g., mediolateral-oblique or MLO view and craniocaudal or CC view mammograms). The three methods are: averaging, taking the maximum, and taking the minimum.

We derived theoretically the performance of each method based on the conventional bivariate binormal model of ROC analysis. In this study, we assumed that both single-view images produce the same ROC curve. The performance of each method depends on the correlation of the diagnostic information derived from the two images and on the single-view ROC curve parameters: a and b. We applied these methods to a mammography study in which the task was to classify clustered microcalcification as malignant or benign based on MLO and CC view mammograms. The database contained 54 benign and 42 malignant cases.

Theoretically, averaging produces improved area under the ROC curve (Az) compared to the single-view images. However, the maximum (when b is small) and the minimum (when b is large) can outperform the average. We constructed a diagram that partitioned the parameter a-b space into three regions that in each region one particular method produced the best performance. Importantly, as the correlation between the two images increases, the region that the average performs best shrinks and the regions that the maximum or the minimum performs best become larger. Results of the mammography study agreed with the theory: both the theory and the experiment found the average to be the best performing method with an Az value of 0.84.

Our theoretical analysis provides insight into the different results that the maximum, the minimum and the average can produce in ROC analysis. Moreover, it provides guidance on selecting an appropriate method to combine diagnostic information obtained from two images of the same patient.

C.E.M.: CEM is a shareholder in R2 Technology, Inc.

Liu, B, Metz, C, Jiang, Y, Effect of Correlation on Combining Diagnostic Information from Two Images of the Same Patient.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4414597.html