RSNA 2003 Scientific Papers > Computerized Regional Registration of Corresponding ...
  Scientific Papers
  SESSION: Physics (Image Processing: CAD I--Breast)

Computerized Regional Registration of Corresponding Masses and Microcalcification Clusters on Temporal Pairs of Mammograms for Interval Change Analysis

  DATE: Monday, December 01 2003
  START TIME: 11:20 AM
  END TIME: 11:27 AM
  CODE: C18-378

Lubomir Hadjiiski PhD
Ann Arbor , MI
Heang-Ping Chan PhD
Berkman Sahiner PhD
Chuan Zhou PhD
Mark Helvie MD
Marilyn Roubidoux MD

Breast neoplasms, calcification
Breast neoplasms, localization
Computers, diagnostic aid

Purpose: To develop a registration technique for automated identification of corresponding lesions on a temporal pair of mammograms of the same view. This technique is the basis for interval change analysis of breast lesions in CAD applications.

Methods and Materials: A multi-stage registration technique is being developed. In the first stage, an initial search region was estimated on the prior mammogram based on the lesion location on the current mammogram. In the second stage the search region was refined. In the third stage the lesion was detected within the search region. For the initial estimation of the lesion centroid location on the prior mammogram, we previously developed a regional registration method (RRM), based on the radial distance between the nipple and the lesion centroid and the angular distance between the nipple-lesion centroid axis and the breast boundary on the current mammogram. In the present study, we compared the RRM to the use of warping techniques for the initial estimation of the lesion location. The current mammogram was warped by affine (AF) or thin plate splines (TPS) transformation in combination with simplex optimization in order to maximize a similarity measure between the breast areas on the current and prior mammograms. Correlation and mutual information (MI) similarity measures were evaluated. A set of 390 temporal pairs of mammograms containing biopsy-proven masses or microcalcification clusters was used. The true lesion locations were identified by an MQSA radiologist on all mammograms. 72 temporal pairs were used for training the parameters of the warping techniques. The remaining 318 pairs were used for testing the performance of the 5 methods. The registration accuracy was analyzed by evaluating the average distance between the centroids of the estimated and the true lesion locations on the prior mammogram.

Results: The average distance between the estimated and the true lesion centroids on the previous mammogram after the initial stage was: RRM = 8.5±6.2mm, correlation-AF = 9.0±6.7 mm, correlation-TPS = 10.3±8.2 mm, MI-AF = 9.2±7.5 mm, MI-TPS = 9.5±8.6 mm. After the final registration stage, the average distance between the estimated and the true centroids was: RRM=6.4±8.9mm, correlation-AF = 7.0±9.5 mm, correlation-TPS = 7.4±10.2 mm, MI-AF = 6.9±9.5 mm, MI-TPS = 7.2±11.1 mm.

Conclusion: The RRM method outperformed the warping techniques. It localized the corresponding lesions on temporal pairs of mammograms with the highest accuracy and the lowest standard deviation among the 5 methods.




Questions about this event email: