Reading Mammography with Multiple Prior Exams

Change is one of the strongest features for identifying abnormality in screening mammography exams. However, publicly available mammography datasets have so far lacked prior exams, and so the majority of algorithm development has focused on assessing isolated exams, without prior information. Recently, it was shown that a deep learning algorithm can improve its diagnostic accuracy by utilizing a single prior mammography image. In this work, we extend the previous result to address the issue of reading a variable number of prior exams. We compare two approaches: Random Forest, which requires that all inputs have the same size, and LSTM, which can handle a variable-size input. We demonstrate a significant performance improvement when using multiple priors, consistent with the standard workflow of breast imagers. We also found for both models that multiple priors improved performance over using a single prior. Interestingly, LSTM consistently outperformed the Random Forest model, and is more practical because it can naturally process any number of prior images that are available at the time of read. We expect that these results will generalize to other screening programs, such as colorectal cancer, where prior images are readily available.