Automated processing of mammograms has been studied for several years. Until the last few years much of the efforts have produced marginal results due to the limitations of inexpensive hardware. Even with the advanced hardware no one has yet taken an entire mammogram and determined whether it is normal or abnormal. We outline a method here to attempt to do just that. As the mammogram is scanned we use image enhancement software to obtain the best contrast and greyscale levels for the image. Because we need to identify both small microcalcifications and relatively large masses we use a multiresolution technique to process the mammograms. We chose the wavelet transform as our multi-resolution tool. Once we have the features from each octave we use them as input to a modular neural network with each octave processed separately at the first stage and then the output brought together at the second stage. We have chosen the ALOPEX training algorithm over backpropagation as it has shown to give better performance in these type of applications.
展开▼
