Impact of multi-source data augmentation on performance of convolutional neural networks for abnormality classification in mammography

Introduction To date, most mammography-related AI models have been trained using either film or digital mammogram datasets with little overlap. We investigated whether not combining and mammography during training will help hinder modern designed for use on mammograms. Methods this end, a total of six binary classifiers were comparison. The first three images only from Emory Breast Imaging Dataset (EMBED) ResNet50, ResNet101, ResNet152 architectures. next EMBED, Curated Subset Digital Database Screening Mammography (CBIS-DDSM), (DDSM) datasets. All tested mammograms EMBED. Results results showed that performance degradation to the customized ResNet was statistically significant overall when EMBED dataset augmented CBIS-DDSM/DDSM. While observed in all racial subgroups, some races are subject more severe drop as compared other races. Discussion may potentially be due ( 1 ) mismatch features between film-based 2 pathologic radiological information. In conclusion, both breast cancer screening. Caution is required utilizing information simultaneously.