Back

Artificial-Intelligence Powered Identification of High-Risk Breast Cancer Subgroups Using Mammography : A Multicenter Study Integrating Automated Brightest Density Measures with Deep Learning Metrics

Jeong, Y.; Lee, J.; Lee, Y.-j.; Hwang, J.; Lee, S. B.; Yoo, T.-K.; Kim, M.-S.; Kim, J. I.; Hopper, J. L.; Nguyen, T. L.; Lee, J. W.; Sung, J.

2024-01-29 radiology and imaging
10.1101/2024.01.28.24301639 medRxiv
Show abstract

Mammography plays a crucial role in breast cancer (BC) risk assessment. Recent breakthroughs show that deep learning (DL) in mammography is expanding from diagnosis to effective risk prediction. Moreover, the brightest mammographic breast density (MBD), termed "cirrocumulus," signifies an authentic risk. Addressing the challenges in quantifying above recent measures, we present MIDAS: a DL-derived system for multi-level MBD and risk feature score (FS). Using >260,000 multicenter images from South Korea and the US, FS consistently outperforms conventional MBD metrics in risk stratification. Only within the high FS, cirrocumulus further enriches assessment, pinpointing "double-higher" subgroup. Their risk profiles are notable: women in double upper-tertile showed OR=10.20 for Koreans and 5.67 for US, and OR=7.09 for scree-detected cases (US only). We also reveals the "black-box" nature of FS that it predominantly captures complex patterns of higher-intensity MBD. Our research enhances the potential of digital mammography in identifying individuals at elevated BC risks.

Matching journals

The top 12 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.