A Data Set and Deep Learning Algorithm for the Detection of Masses and Architectural Distortions in Digital Breast Tomosynthesis Images

Authors:
Top Cited Papers
Publisher website
Google Scholar
Abstract
Deep learning emerged mainly due to rapid increases in access to computational resources and large-scale labeled data.1 Medical imaging is a natural application of deep learning algorithms.2 However, well-curated data are scarce, which poses a challenge in training and validating deep learning models. Annotated medical data are limited for a number of reasons. First, the number of available medical images is much lower than the number of available natural images. This is particularly an issue when investigating a condition with fairly low prevalence, such as breast cancer in a screening setting (<1% of screening examinations result in a cancer diagnosis). Second, access to medical imaging data is guided by a number of strict policies given that they contain patients’ medical information. Sharing of medical imaging data requires an often nontrivial and time-consuming effort to deidentify the data as well as ensure compliance with requirements from the institution that is sharing the data and beyond. Finally, annotation of medical imaging data typically requires the work of radiologists, who already have high demands on their time.

This publication has 15 references indexed in Scilit: