Prediction of Near-Term Breast Cancer Occurrence using Subtraction of Temporally Sequential Digital Mammograms.

Breast cancer remains one of the leading cancers for women worldwide. Fortunately, with the introduction of mammography, the mortality rate has significantly decreased. However, earlier breast cancer prediction could effectively increase the survival rates, improve patient outcomes, and avoid unnecessary biopsies. For that purpose, prediction of breast cancer, using subtraction of temporally sequential digital mammograms and machine learning, is proposed. A new dataset was collected with 192 images from 32 patients (three screening rounds, with two views of each breast). This dataset included precise annotation of each individual malignant mass, present in the most recent mammogram, with the two priors being radiologically evaluated as normal. The most recent mammogram was considered as the "future" screening round and provided the location of the mass as the ground truth for the training. The two previous mammograms, the "current" and the "prior", were processed and a new, difference image was formed for the prediction. Ninety-six features were extracted and five feature selection algorithms were combined to identify the most important features. Ten classifiers were tested in leave-one-patient-out and k-fold-patient cross-validation (k = 4 and 8). Ensemble Voting achieved the highest performance in the prediction of the development of breast mass in the next screening round, with 85.7% sensitivity, 83.7% specificity, 83.7% accuracy and 0.85 AUC. The proposed methodology could lead to a new mammography-based model that could predict the short-term risk for developing a malignancy, thus providing an earlier diagnosis.