Towards an Unsupervised GrowCut Algorithm for Mammography Segmentation

Breast cancer is the most frequent type of malignancy in women, with 2.3 million diagnostics only in 2020. However, as a consequence of early diagnosis and appropriate treatment, more and more women are being cured. Among screening methods, mammography is one of the most used, and segmentation is a crucial step of its analysis. Starting from an enhanced version of the GrowCut segmentation technique, the goal of this paper is to automatically generate the foreground seeds so as to minimize the human expert intervention to the identification of only one pixel. We propose a method that starts with the center of the anomaly and composes the foreground seeds set as a circle inside the anomaly. We test the proposed method on the mini-MIAS dataset for various radius dimensions and two background seeds variants, concluding that the best combination is a circle with a radius of 25 pixels as foreground seeds and the black pixels from the original image as background seeds.