Accurate lesion segmentation in mammography is important for computer-aided breast cancer diagnosis, but it typically relies on dense pixel-level annotations. These annotations are costly to obtain and limited in availability, which constrains the development of robust models.

The paper, posted on arXiv, introduces a lesion-aware attention mechanism designed specifically for mammography segmentation. The method is intended to work with heterogeneous supervision—meaning it can use different types or levels of annotation rather than requiring full pixel-level labels for every training image. This is a practical advantage because existing public segmentation datasets, as the abstract notes, usually cover only a small set of examples.

By combining lesion-aware attention with heterogeneous supervision, the work aims to reduce the annotation burden while still supporting accurate segmentation. The approach is relevant to computer-aided diagnosis systems, where labeled medical imaging data is often scarce and expensive to produce.