There is only a small subset of tumors that you can identify from mammography. Many tumors do not show up at all in X-ray mammography images, while others may look suspicious but not end up being malignant lesions. So, identification of of tumors on mammo images is not really a software issue, but a hard, probably unsolvable clinical problem (using mammography alone is just not enough to identify tumors).
Of course you can still train a neural network to identify suspicious lesions on mammo images, and even though it is not perfect, can still be practically useful. I would assume that there are lots of training data sets available to train your network and probably there are dozens of solutions already proposed for this. So, I would recommend to start with some web and literature search and try to reproduce results reported by others.