Y. Sun, J. Suri (USA), R.M. Rangayyan (Canada), and R. Janer (USA)
Screening mammogram, breast skin-line, performance evaluation, Hausdorff distance
: Performance evaluation of breast skin-line estimation algorithms is a crucial step for standardization of Computer-Aided Detection (CAD) techniques applied to mammograms. A good quantitative analysis for skin line will benefit and facilitate (a) breast region segmentation algorithms and (b) the digital image acquisition chain including digital detectors. However, there has been no consensus on the metric to be used for evaluation of the skin-line boundaries. This paper presents a close look at the metrics used for error measurement between the ground truth boundaries (traced by radiologists) and the automatic computer estimated boundaries of breast skin-lines. We demonstrate the comparison of two major metrics: Suri's Polyline Distance Measure (PDM) [1-4] and the Hausdorff Distance Measure (HDM) based on a distance transform and image algebra [5]. In addition, we also present a system which automatically (a) estimates normalized False Negative Fraction (FNF) and False Positive Fraction (FPF) measures, and (b) spots out those skin-line boundaries which are not within a radiologist's accepted error threshold based on quartile measurement. Our error metric techniques were applied to 83 images from the MIAS database [6], where the computer estimated boundaries were computed using the Deformable Model developed by Ferrari et al. [7]. The PDM method yielded a mean error () of 2.49 pixels with a standard deviation () of 3.69 pixels. The HDM method yielded of 21.06 pixels and of 10.56 pixels. The normalized FNF was 0.57% and the normalized FPF was 1.27%.
Important Links:
Go Back
