Microcalcification Detection in Mammograms Using Difference of Gaussians Filters and a Hybrid Feedforward-Kohonen Neural Network

This work develops a microcalcifications´ detection system in mammograms by using difference of Gaussians filters (DoG) and artificial neural networks (ANN). The digital image processing proposed show the basic wavelet-based behavior of DoG as a mother function frequently used in several vision tasks, and in this case, used in order to enhance the microcalcifications´ traces in standard mammograms and further to achieve its detection via ANN. In order to achieve this, a segmentation algorithm is implemented for reaching a threshold in already processed images, and finally, the resultant information is given to the ANN. The neural network used to perform the detection is a hybrid feedforward-Kohonen one, implemented with a hard-limit transfer function in the first layer and a self-organizing map (SOM) responsible for microcalcifications´ topologic adjustment in the second layer. Basically, this clustering method gave us a robust solution of the problem and the detection was performed efficiently. There are no considerations relative to morphologic analysis of microcalcifications for diagnosis in this work.