Deep-LIBRA: An artificial-intelligence method for robust quantification of breast density with independent validation in breast cancer risk assessment

Abstract

Breast density is an important risk factor for breast cancer that also affects the specificity and sensitivity of screening mammography. Current federal legislation mandates reporting of breast density for all women undergoing breast cancer screening. Clinically, breast density is assessed visually using the American College of Radiology Breast Imaging Reporting And Data System (BI-RADS) scale. Here, we introduce an artificial intelligence (AI) method to estimate breast density from digital mammograms. Our method leverages deep learning using two convolutional neural network architectures to accurately segment the breast area. An AI algorithm combining superpixel generation and radiomic machine learning is then applied to differentiate dense from non-dense tissue regions within the breast, from which breast density is estimated. Our method was trained and validated on a multi-racial, multi-institutional dataset of 15,661 images (4,437 women), and then tested on an independent matched case-control dataset of 6368 digital mammograms (414 cases; 1178 controls) for both breast density estimation and case-control discrimination. On the independent dataset, breast percent density (PD) estimates from Deep-LIBRA and an expert reader were strongly correlated (Spearman correlation coefficient = 0.90). Moreover, in a model adjusted for age and BMI, Deep-LIBRA yielded a higher case-control discrimination performance (area under the ROC curve, AUC = 0.612 [95% confidence interval (CI): 0.584, 0.640]) compared to four other widely-used research and commercial breast density assessment methods (AUCs = 0.528 to 0.599). Our results suggest a strong agreement of breast density estimates between Deep-LIBRA and gold-standard assessment by an expert reader, as well as improved performance in breast cancer risk assessment over state-of-the-art open-source and commercial methods.

Keywords: Artificial intelligence; Breast cancer risk; Breast density; Deep learning; Digital mammography.

Fig. 1:

Development and evaluation experiments. White boxes: workflow of the Deep-LIBRA algorithm. Green, blue, yellow, and red boxes: training, validation, independent testing, and blinded independent testing, respectively. HUP: Hospital of the University of Pennsylvania; MC: Mayo Clinic.

Fig. 2:

Detailed illustration of the Deep-LIBRA algorithm operation. Panel (a) shows the original FFDM image in 16-bit resolution, and panel (b) is the zero-padded image in an 8-bit intensity resolution. The zero-padded image is used by the background segmentation U-Net, which generates the image shown in panel (c). Panel (d) is the output of the module of pectoralis muscle removal using the second U-Net resulting to the final breast segmentation shown in panel (e). The image from panel (e) is used to generate superpixels as shown in panel (f) and perform radiomic feature analysis. Finally, the SVM classifies the superpixels based on the extracted features, resulting in dense tissue segmentation, as shown in panel (g). The panel (h) shows the final dense tissue segmentation overlaid on the original image. Note: The image sizes are different in this figure because the panels (a), (e)-(h) show images in the original image resolution, while the panels (b)-(d) are downsampled images of size 512 × 512 pixels used in U-Net segmentation.

Fig. 3:

The majority voting approach. The majority voting approach uses the outcome of three SVM models, each trained on two folds of ds3-a, to make the final dense tissue segmentation. The majority voting scheme assigns the dense or non-dense label to each superpixel based on at least two SVM models agreeing on the label.

Fig. 4:

Deep-LIBRA evaluation curves in the development phase. Panels (a) and (b) show the training and validation (noted as as “val_”) results for background and pectoral muscle segmentation CNNs, respectively. As the panel (b) shows, there is no sign of overfitting for pectoralis muscle segmentation while panel (a) indicates some possible signs of overfitting after epoch 40 shown by a wider fluctuation on the validation set.