Publication

Abstract : Breast cancer, a potentially fatal condition, affects one in twenty-eight women in India. Breast density, which is defined as the ratio of the volume of breast tissue to that of fatty tissue, is an important criterion in measuring the likelihood of a person developing breast cancer. In breast cancer, it is almost always the breast tissue that turns cancerous, not the fatty tissue. Women who have dense breast tissue, are four to six times more likely to develop breast cancer than women who do not have dense breast tissue. In our study, we examine breast mammograms and aim to develop a technique to automatically quantify breast density objectively, which is important in the prediction of the risk of breast cancer. While increased breast density greatly predisposes the risk of breast cancer, it is important to assess the risk quantitatively, which is difficult. In our technique, we develop a technique based on U-Net to segment breast tissues. We aim to classify the breasts based on the amount of breast tissue they have compared with the total volume of the breast. Therefore, we utilize a training database of 20 human-segmented images to train our algorithm and test it on a separate set of 15 human-segmented images. Examining the accuracy of the segmentation, we find that we obtain an accuracy of 0.79±0.12 in the Dice similarity coefficient. The qualitative images have also been examined and as may be observed from the qualitative results. A fully replicable, open-source code that everyone can use to segment breast tissue from mammogram images is the project's final goal.