A Multisite Study of a Breast Density Deep Learning Model for Full-Field Digital Mammography and Synthetic Mammography

Abstract

B reast density is an important risk factor for breast cancer (1-3). Additionally, areas of higher density can mask findings within mammograms, leading to lower sensitivity (4). Many states have passed breast density notification laws requiring clinics to inform women of their breast density (5). Radiologists typically assess breast density by using the Breast Imaging Reporting and Data System (BI-RADS) lexicon, which divides breast density into four categories: A, almost entirely fatty; B, scattered areas of fibroglandular density; C, heterogeneously dense; and D, extremely dense (examples are presented in Fig E1 [supplement]) (6). Unfortunately, radiologists exhibit intra-and interreader variability in the assessment of BI-RADS breast density, which can result in differences in clinical care and estimated risk (7-9).Deep learning (DL) has previously been used to assess BI-RADS breast density for film (10) and full-field digital mammographic (FFDM) images (11-16), with some models demonstrating closer agreement with consensus estimates than individual radiologists (14). To realize the promise of the use of these DL models in clinical practice, two key challenges must be met. First, because digital breast tomosynthesis (DBT) is increasingly used in breast cancer screening (17) due to improved reader performance (18-20), DL models should be compatible with DBT examinations. To aid in radiologist interpretation of breast cancer and breast density, DBT examinations contain twodimensional images in addition to three-dimensional images. These two-dimensional images may be either FFDM images or synthetic two-dimensional mammographic (SM) images derived from the three-dimensional images. Figure E2 (supplement) shows the differences in image characteristics between FFDM and SM images. The relatively recent adoption of DBT at many institutions means that the datasets available for training DL models are often fairly limited for DBT examinations compared with FFDM examinations. Second, DL models must offer consistent performance across sites, where differences in imaging technology, patient demographics, or assessment practices could impact model performance. To be practical, this

Artificial intelligence in mammographic phenotyping of breast cancer risk: a narrative review

“…Most recently, in what could be an essential step towards AI-enabled BI-RADS breast density assessment, research has focused on leveraging domain adaptation approaches to create DL models that utilize 2D synthetic mammographic (SM) images reconstructed from DBT acquisitions. The feasibility of this approach was demonstrated in large, racially diverse datasets from two clinical sites, where the adapted model achieved good agreement with the BI-RADS density classification from SM images by radiologists (four-class K = 0.72–0.79) [ 26 ]. Additional novel directions in this field include exploring state-of-the-art DL architectures [ 27 , 28 ], as well as using federated learning, where participating institutions share model weights amongst themselves instead of the actual images.…”

“…As of now, breast density evaluation from 3D reconstructed image volumes has only been explored via traditional radiomic AI techniques [ 40 ] and no DL models have been extended to 3D DBT images. Moreover, with a few exceptions [ 26 , 30 , 38 ], most DL models have been developed using racially homogeneous datasets of FFDM images from a single vendor acquired at a single site, which may limit their ability to generalize to diverse breast cancer screening populations.…”

“…Furthermore, due to a lack of large 3D image datasets, 3D DL models usually need to be trained from scratch, which can lead to unstable convergence and poor generalization issues. Therefore, extensive work is needed to develop DL architectures which are suitable for DBT, as well as to determine whether the knowledge, training data, and models developed for FFDM can be applied to DBT [ 26 ].…”

When Doctors and AI Interact: on Human Responsibility for Artificial Risks

“…To tackle this issue, Matthews et al, (2020) proposed a DL system for breast density classification. The main difference between this DL-based experiment with the previous SVM-based proposal lies in the methodology followed by the AI system designers to enable it to perform classification of medical images.…”

A deep learning system to obtain the optimal parameters for a threshold-based breast and dense tissue segmentation

“…, or pattern recognition based classification tasks, using fully connected convolutional neural networks[48,49], or Deep Residual Learning for BI-RADS breast density categories classification[50]. Since our output was continuous, approaches intended to pixel-level classification were discarded.…”