Breast cancer is the most commonly diagnosed cancer in women. The key for its early detection is regular screening. One of the most widely used methods for early breast cancer detection is mammography. For decades, computer aided diagnostic systems have been utilized to assist physicians in the analysis of mammograms.
In my thesis I examined the possibilities of applying deep learning methods in breast imaging, which have achieved great success in the field of computer vision. Using a dataset of cropped mammographic images I designed deep neural network models which achieved high accuracy on the classification of these images. I then utilized the networks as a part of an analyzer program which can examine full scale mammograms.
On the cropped mammographic image set, the created models achieved 96% classification accuracy on microcalcifications and 93% on masses. As the part of the full scale mammogram analyzer program, an ensemble of models reached around 96% sensitivity while detecting microcalcifications, averaging 2,56 false positives per image. The same results for the detection of masses was 79% sensitivity with 1,95 false positives per image.