Microcalcification detection in mammography

End-to-end deep learning detection of microcalcifications using a U-Net architecture, with downstream localization of asymmetric patterns. Peer-reviewed in Patterns (Cell Press).

Peer-reviewed · Patterns (Cell Press), 2020

Idea

Microcalcifications are among the earliest visible markers of breast cancer on a mammogram, but they are tiny — often a few pixels wide — and easy to miss against the surrounding tissue. The goal here was end-to-end detection: a deep learning model that segments microcalcifications directly from the raw mammogram, no hand-crafted features or multi-stage pipelines.

Approach

A U-Net segmentation backbone trained on annotated mammograms, with a detection head producing pixel-level masks. The downstream pipeline locates asymmetric patterns and surfaces candidate cancer regions for radiologist review. Performance was evaluated on held-out studies and compared against the standard CAD baselines of the time.

Outcome

Published as part of “Detecting asymmetric patterns and localizing cancers on mammograms” in Patterns (Cell Press), 2020 — Guan, Wang, Li, Zhang, Chen, Siddiqui, Nehring, Huang.

Read the paper →