SU‐E‐I‐05: A Novel Application of SIFT Technique to Microcalcification Detection

Abstract

Purpose: This work applies Lowe's Scale Invariant Feature Transform (SIFT) to detect micro‐calcification on mammograms. The objective of this study is to expand the function of SIFT, which has originally been used to match objects by matching the detected feature points, and to provide a new method for micro‐calcification detection. Methods: First, variables in SIFT, the scaling factor between levels of the image, the radiuses of the areas for maximum comparison within current scale and neighboring scales, and the threshold value for maximum search, were adjusted to allow nearly all the micro‐calcification to be detected as the feature points. Second, to reject feature points which are not micro‐calcification, four features of sixty‐five feature points, determined by physicians as micro‐calcification, curvature of scale space, elements of Hessian matrix, used for the discrimination of prominence and shapes, and Contrast to neighboring pixels, Size on image used to reject points on blob‐like dense tissue, were analyzed to determine the specific ranges for selecting feature points on micro‐calcification. Results: Ninety region of interest (ROI) images (268 × 268 pixel) selected from 85 mammograms (3,540 × 4,740 pixel) were employed to test the proposed method. Of the 90 ROI images, 30 images were biopsy‐verified by a physician to present a cluster of micro‐calcification. The other 60 ROI images are selected from normal mammograms. The performance of the study is evaluated by a receiver operating characteristic curve (ROC). An area under the ROC curve of 94.2%, sensitivity of 93.3%, and specificity of 95% was achieved. Conclusion: The proposed system based on SIFT accurately detects micro‐calcification in mammograms with various brightness, size, and breast density, without preprocessing them. Our findings merit further investigation for its potential to classify benign and malignant calcification.