Detection of microcalcification with neural networks

We propose a detection system for microcalcifications in digitized mammograms. From the MIAS database we selected a few images. From these images the coordinates of some of the calcified points (positive points) are manually identified and a set of normal points (negative points) are randomly selected. Then a set of 87 features is calculated for each of the selected pixels. A neural network based online feature selection method is used to identify a set of good features from these features. We trained a multilayered perceptron (MLP) with the selected features. The false positive output of the network is reduced using connected component analysis. Linear or curved structures are then removed to get a further cleaned image. We then applied the mountain clustering algorithm on the cleaned image to find out dense regions as suspected areas. The proposed system not only is able to classify an image as normal or abnormal, but also for an abnormal image it points out the suspected area with microcalcifications. The system is tested on several images and obtained good results.