Model uncertainty estimates for deep learning mammographic density prediction using ordinal and classification approaches.
Authors
Affiliations (5)
Affiliations (5)
- University of Exeter, Exeter, Exeter, England, EX4 4QJ, UNITED KINGDOM OF GREAT BRITAIN AND NORTHERN IRELAND.
- University of Toronto, Toronto, Toronto, M5S 1A1, CANADA.
- Division of Evolution, Infection and Genomics, School of Biological Sciences, The University of Manchester, Oxford Road, Manchester, M13 9PL, UNITED KINGDOM OF GREAT BRITAIN AND NORTHERN IRELAND.
- Sunnybrook Health Sciences Centre, Toronto, Toronto, Ontario, M4N 3M5, CANADA.
- University of Manchester - The Victoria University of Manchester Campus, Stopford Building, Oxford Road, MANCHESTER, M13 9PT, Manchester, England, M13 9PL, UNITED KINGDOM OF GREAT BRITAIN AND NORTHERN IRELAND.
Abstract
Mammographic density is associated with the risk of developing breast cancer and can be predicted using deep learning methods. Model uncertainty estimates are not produced by standard regression approaches but would be valuable for clinical and research purposes. Our objective is to produce deep learning models with in-built uncertainty estimates without degrading predictive performance.

Approach: We analysed data from over 150,000 mammogram images with associated continuous density scores from expert readers in the Predicting Risk Of
Cancer At Screening (PROCAS) study. We re-designated the continuous density scores to 100 density classes then trained classification and ordinal deep learning models. Distributions and distribution-free methods were applied to extract predictions and uncertainties. A deep learning regression model was trained on the continuous density scores to act as a direct comparison.

Results: The root mean squared error (RMSE) between expert assigned density labels and predictions of the standard regression model were 8.42 (8.34-8.51) while
the RMSE for the classification and ordinal classification were 8.37 (8.28-8.46) and 8.44 (8.35-8.53) respectively. The average uncertainties produced by the models were higher when the density scores from pairs of expert readers density scores differ more, when different mammogram views of the same views are more variable, and when two separately trained models show higher variation.

Conclusions: Using either a classification or ordinal approach we can produce model uncertainty estimates without loss of predictive performance.