Improvement of a Radiomics Based Automated Breast Density Algorithm Evaluated on a Time Series of Mammograms

Radiographic breast density is an important independent breast cancer risk predictor. Current automated density estimation algorithms have a limited prediction power. Normally they analyze full field digital mammograms recorded at a single screening visit. We have developed innovative time-series image analytics and compared it to a single time-point prediction.

We looked at 3200 cancer free participants of the Slovenian Breast Cancer Screening Programme DORA that had a mammogram that needed further screening assessment. Our prediction algorithm was optimized to the breast density scored by a radiologist as part of the assessment. We combined multiple predictions from the series using a soft voting classifier with specific weights optimized for accuracy.

For the reference visit only, the Cohen kappa score was 0.63±0.01. Pairing them with earlier predictions, kappa grew to 0.67±0.01, and earlier predictions were favored with a weight of 0.55±0.05. With future predictions, no improvements were seen, and low weights were assigned.

The accuracy of our density scoring algorithm with performance comparable to algorithms reported in literature, was improved slightly using a time-series of mammograms. The improvement relied on earlier, but not on mammograms recorded after the scoring visit.