Segmenting mammographic microcalcifications using a semi-automatic procedure based on Otsu's method and morphological filters

ABSTRACT

INTRODUCTION: Breast cancer has the second highest world's incidence rate, according to the Brazilian National Cancer Institute (INCa). Clinical examination and mammography are the best methods for early diagnosis. Computer-aided detection (CADe) and computer-aided diagnosis (CADx) systems are developed to improve mammographic diagnosis. Basically, CADx systems have three components (i) segmentation, (ii) parameters extraction and selection, (iii) lesion classification. The first step for a CADx system is segmentation. METHODS: A microcalcification segmentation method is proposed, based on morphological operators, Otsu's Method and radiologists' knowledge. Pre-processing with top-hat operators improves contrast and reduces background noise. The Otsu's method automatically selects the best grey-level threshold to segment microcalcifications, obtaining binary images. Following, inferior reconstruction and morphological dilatation operators are applied to reconstruct lost structure details and fill small flaws in the segmented microcalcifications. Finally, the Canny edge detection is applied to identify microcalcifications contour candidates for each region-of-interest (ROI). Two experienced radiologists intervene in this semi-automatic method, firstly, selecting the ROI and, then, analyzing the segmentation result. The method was assessed in 1000 ROIs from 158 digital images (300 dpi, 8 bits). RESULTS: Considering the radiologists opinion, the rates of ROIs adequately segmented to establish a diagnosis hypothesis were 97.8% for one radiologist and 97.3% for the other. Using the Area Overlap Measure (AOM) and the 2136 microcalcifications delineated by an experienced radiologist as gold standards, the method achieved an average AOM of 0.64±0.14, being 0.56±0.09 for small microcalcifications and 0.66±0.13 for the large ones. Moreover, AOM was 0.64±0.13 for the benign and 0.64±0.14 for the malignant lesions with no statistical differences between them. CONCLUSION: Based on these findings, the proposed method could be used to develop a CADx system that could help early breast cancer detection.