Detection and quantification of breast arterial calcifications on mammograms: a deep learning approach

Abstract

Objective: Breast arterial calcifications (BAC) are a sex-specific cardiovascular disease biomarker that might improve cardiovascular risk stratification in women. We implemented a deep convolutional neural network for automatic BAC detection and quantification.

Methods: In this retrospective study, four readers labelled four-view mammograms as BAC positive (BAC+) or BAC negative (BAC-) at image level. Starting from a pretrained VGG16 model, we trained a convolutional neural network to discriminate BAC+ and BAC- mammograms. Accuracy, F1 score, and area under the receiver operating characteristic curve (AUC-ROC) were used to assess the diagnostic performance. Predictions of calcified areas were generated using the generalized gradient-weighted class activation mapping (Grad-CAM++) method, and their correlation with manual measurement of BAC length in a subset of cases was assessed using Spearman ρ.

Results: A total 1493 women (198 BAC+) with a median age of 59 years (interquartile range 52-68) were included and partitioned in a training set of 410 cases (1640 views, 398 BAC+), validation set of 222 cases (888 views, 89 BAC+), and test set of 229 cases (916 views, 94 BAC+). The accuracy, F1 score, and AUC-ROC were 0.94, 0.86, and 0.98 in the training set; 0.96, 0.74, and 0.96 in the validation set; and 0.97, 0.80, and 0.95 in the test set, respectively. In 112 analyzed views, the Grad-CAM++ predictions displayed a strong correlation with BAC measured length (ρ = 0.88, p < 0.001).

Conclusion: Our model showed promising performances in BAC detection and in quantification of BAC burden, showing a strong correlation with manual measurements.

Clinical relevance statement: Integrating our model to clinical practice could improve BAC reporting without increasing clinical workload, facilitating large-scale studies on the impact of BAC as a biomarker of cardiovascular risk, raising awareness on women's cardiovascular health, and leveraging mammographic screening.

Key points: • We implemented a deep convolutional neural network (CNN) for BAC detection and quantification. • Our CNN had an area under the receiving operator curve of 0.95 for BAC detection in the test set composed of 916 views, 94 of which were BAC+ . • Furthermore, our CNN showed a strong correlation with manual BAC measurements (ρ = 0.88) in a set of 112 views.

Keywords: Artificial intelligence; Cardiovascular diseases; Mammography; Risk factors; Vascular calcification.

Fig. 1

Examples of breast arterial calcifications on screening mammograms (white arrows). a Low, b mild, and c severe burden of BAC

Fig. 2

General VGG16 architecture consisting of 13 convolutional layers (kernel 3 × 3, depth k), 5 pooling layers (non-trainable), and 2 fully connected (FC, n: number of neurons) layers followed by a Softmax activation function to solve the multiclass classification problem (a), and the final CNN for automated binary BAC detection where the “non-trainable layers” exploited VGG16 transfer learning (b). Rectified linear unit (ReLU) activation functions (in model a) and leaky ReLUs (in model b) following each convolutional kernel are not shown

Fig. 3

ROC and PR curves of training (red line), validation (blue line), and test (green line) subsets

Fig. 4

Visual explanations (Grad-CAM++ heatmaps) of the automatic detection results by the proposed model. a, a′ True-positive case with a high burden of BAC in multiple vessels; b, b′ true-positive case with small BAC (arrows); c, c′ true-negative case with confounding factors, i.e. various benign calcifications (none of the structures colored on the heatmap reaches the threshold for being finally detected as BAC)

Fig. 5

Examples of misclassification. a, a′ False-positive case with small calcifications within a Cooper’s ligament mistaken as BAC (arrow), b, b′ false-positive case with skinfold including cutaneous calcifications mislabelled as BAC (arrowhead), c, c′ false-negative case with small BAC concealed under dense tissue (circle)

Fig. 6

a Automatic segmentation of a BAC by thresholding the Grad-CAM++ heatmap of a mammogram with moderate burden of BAC (length 41 mm). b Scatterplot of the estimated area (y-axis) compared to the manually measured length (x-axis) for all 56 women in the subgroup (112 views)