RSNA 2003 Scientific Papers > Observer Evaluation and CAD Performance of a Radial ...
  Scientific Papers
  SESSION: Physics (Image Processing: CAD I--Breast)

Observer Evaluation and CAD Performance of a Radial Gradient-based Segmentation Method for Mammographic Microcalcifications

  DATE: Monday, December 01 2003
  START TIME: 11:30 AM
  END TIME: 11:37 AM
  CODE: C18-379

Sophie Paquerault PhD
Chicago , IL
Laura Yarusso
Robert Nishikawa PhD
John Papaioannou MS
Alexandra Edwards
Yulei Jiang PhD

Breast, calcification
Computers, diagnostic aid
Images, processing


Purpose: To perform an observer evaluation of the accuracy of a proposed segmentation method for microcalcifications in mammograms and to demonstrate the improvement in performance of our computerized classification scheme for malignant versus benign microcalcifications when using the proposed segmentation technique.

Methods and Materials: We have implemented a radial gradient-based segmentation method for microcalcifications. It is difficult task for radiologists to manually outline the contours of each microcalcification. Therefore, we have conducted two observer studies to evaluate the proposed segmentation method. Two observers participated in each study, reviewing a database of 50 digitized mammograms with clustered microcalcifications. No observer participated in both studies. The first observer study (A) required each observer to rate the accuracies (0-100 scale) of both the proposed segmentation method and the one currently used in our computerized classification scheme. In the second observer study (B), each observer was asked to select their preferred method from three displayed segmentation methods: the proposed segmentation method, the watershed method and the current method. In these two studies, the original mammograms with no segmentation were also displayed. In addition, the 50 cases were randomly reviewed 2 to 3 times to analyze observers' consistencies. The effect of the proposed segmentation method on the performance of our computerized classification scheme was also investigated. ROC analysis was performed and the Az value was used as a performance index.

Results: For study (A), the first observer gave accuracy ratings of 83 and 31 for the proposed and current segmentation methods respectively. The second observer also preferred the proposed method, with accuracy ratings of 92 versus 68 for the current method. For study (B), the observers selected the proposed method among the three displayed methods 56% and 62% of the time. Their other selections were equally distributed between the other two segmentation methods. The performance of the classification scheme improved when using the proposed segmentation method. The Az values for case-based performance were 0.84 with the proposed method versus 0.78 with the current method.

Conclusion: Two separate observer studies demonstrated that our proposed radial gradient-based segmentation technique for mammographic microcalcifications was preferred over other methods. The proposed method also improved the performance of our CAD classification scheme. (R.M.N., J.P. are shareholders in R2 Technology.)

Questions about this event email: