Breast cancer can be screened, diagnosed, and managed earlier with digital mammograms. Data shortages are one of the main challenges in many medical image processing tasks, including this one. There are only a limited number of cases with reliable labeling and accurate marking of the tumor regions. On the other hand, images taken by different mammography devices have different quality, brightness, etc. (generally called domain), depending on the type, brand, and technology used in the device. Therefore, a model trained on a dataset, will not necessarily have a good performance on another one. In developing commercial models for the processing of mammography images, domain adaptation is thus an essential, and yet complicated issue.
This project aims to extend the accuracy produced on a mammography screening dataset with accurate labeling and marking to an acceptable level on other datasets with different domains.