Outlier Detection for Mammograms

Ryan Zurrin, Neha Goyal, Pablo Bendiksen, Muskaan Manocha, Dan Simovici, Nurit Haspel, Marc Pomplun, Daniel Haehn

Keywords: anomaly detection, outlier detection, mammograms, unsupervised learning

TL;DR: We identify low-quality mammograms using a combination of min-max normalized histogram binning and a variational autoencoder in a two-stage automated pipeline.

Abstract: Mammograms are vital for detecting breast cancer, the most common cancer among women in the US. However, low-quality scans and imaging artifacts can compromise their efficacy. We introduce an automated pipeline to filter low-quality mammograms from large datasets. Our initial dataset of 176,492 mammograms contained an estimated 5.5% lower quality scans with issues like metal coil frames, wire clamps, and breast implants. Manually removing these images is tedious and error-prone. Our two-stage process first uses threshold-based 5-bin histogram filtering to eliminate undesirable images, followed by a variational autoencoder to remove remaining low-quality scans. Our method detects such scans with an F1 Score of 0.8862 and preserves 163,568 high-quality mammograms. We provide results and tools publicly available as open-source software.