Adaptive window and adaptive threshold method for microcalcification detection

Microcalcifications are identified in mammograms by pixels that have a brightness greater than their boundary. The distribution of detected microcalcifications depends on their intensity and window size. We propose an adaptive window and adaptive threshold (AWAT) method for microcalcification detection. Based on the intensity of the distribution, we found that 2 times of standard deviation (2SD) is the optimum threshold for detecting local maxima. We also propose an adaptive window that is dependent on the surrounding tissue. The local maxima are identified using a threshold adapted to each window. Then, small objects are removed using morphological operations. The remaining local maxima are called candidates and classified into microcalcification or normal tissue using three classifiers: a multilayer perceptron, a radial basis function neural network, and a support vector machine. We compared the results of our method to a method using five different fixed window sizes, evaluating the performance using the area under the receiver operating characteristic curve. Our experimental results revealed that our method outperformed all the fixed window size approaches, for all the classifiers investigated. Overall, the multilayer perceptron performed the best among the classifiers, with area under ROC curve A_z=0.951 (compared with A_z=0.916 and A_z=0.847). Finally, the results found that size of window varies from 10 and 131 pixels, while the threshold also varies from 5.628 and 229.959 of intensity. In the spatial domain, both threshold and window size are required to detect local maxima. The results of our experiments demonstrated that the proposed AWAT method performed better than fixed window size methods.