Breast cancer is the second leading cause of cancer deaths among U.S. women. Screening mammography has been found to reduce mortality rates, but also comes with a high risk of false positives and false negatives. Detection of breast cancer on screening mammography is challenging as an image classification task because cancerous tissue only represents a small portion of the tissue in the image.
Deep learning and transfer learning have proven to be very effective at improving the precision and recall of automated detection of breast cancer in mammography screening images. Starting from a pre-trained model allows Retina users to train an effective classifier from a minimal amount of annotated images in a relatively short amount of time. This sort of model can help triage cases, helping radiologists prioritize the most suspicious cases.
Models trained by Retina achieved state-of-the-art results on the Breast Cancer Classification from Histopathological Images dataset, a well known, publicly available dataset of 7909 images of either benign or malignant tumors.