Deep Transfer Learning to Classify Mass and Calcification Pathologies from Screen Film Mammograms

Abstract

The number of breast cancer diagnoses is the largest among all cancers among women in the world. Breast cancer treatment is possible if it is diagnosed in the early stages. Mammography is a common imaging technique to detect breast cancer abnormalities. Breast cancer symptom screening is being performed by radiologists. In the last decade, deep learning was successfully applied to big image classification databases such as the ImageNet. In this study, the breast cancer pathology classification performances of the recent deep learning models were investigated by transfer learning and fine tuning. A total of 3,360 mammogram patches were used from the Digital Database for Screening Mammography (DDSM) and the Curated Breast Imaging Subset of DDSM (CBIS-DDSM) mammogram databases for deep learning model training, validating, and testing. Transfer learning and fine tuning were applied using Resnet50, Xception, NASNet, and EfficientNet-B7 network weights. The best classification performance was achieved by transfer learning from the Xception network. The computational costs of deep learning models were considered while selecting the best one. On the original CBIS-DDSM five-way test mammogram classification problem, the mean sensitivity, specificity, F1-score, and AUC were 0.7054, 0.9264, 0.7024, and 0.9317, respectively. The results show that the proposed models may be useful for the classification of breast cancer pathologies.