Detection of Microcalcifications in Mammograms Using Algorithmic Filtering and Region-Based Neural Networks
Dániel Hadházi, Mihály Vetró, Gábor Hullám · 2024
In digital mammography, which is a common method for breast cancer screening in women, the presence of small areas of calcified tissue, called microcalcifications, is often indicative of rapid tissue growth and therefore serves as one of the most common early signs of breast cancer. The small size of these microcalcifications (typically having a diameter of a fraction of a millimeter) makes their detection difficult even for experienced radiologists, making the development of computer-aided detection (CADe) methods critical. In this paper, we present a novel method based on the combination of algorithmic filtering and region-based neural networks for the automated detection of microcalcifications in mammograms. This method utilizes a constrained h-maxima transform based filtering for the enhancement of calcification like regions. After the filtering, vascular calcification patterns are detected and eliminated by the cascade of algorithmic filtering and a region based convultional neural network (R-CNN). An adaptive thresholding is utilized for the detection of final microcalcification region of interests (ROIs), which are discriminated by a second convolutional neural network. The proposed method is able to achieve 48% sensitivity with less than 1 average false positive per image on the INbreast digital mammography database, which in some cases contains more than 100 individual calcified regions within a single mammogram.