Classifying the Presence or Absence of Calcifications on Mammography using the EfficientNet: A Feasibility Study

Calcifications on mammography are important to detect malignant breast lesions. Here a deep learning-based classifier is used to differentiate the presence and absence of breast calcifications on mammography images. A total of 1471 calcification lesions included in the Digital Database for Screening Mammography were used for the dataset. High-density breasts (breast density of ≥4) were excluded from the dataset. The calcification lesion dataset was created from randomly cropped patch images from each lesion that overlapped calcification lesion masks. The non-calcification dataset was created using sliding window patch images and excluded any calcification lesion mask overlap. Subsequently, the same number of patch images was created. MLO and CC binary classifiers were created using the EfficientNet(B0). The evaluation was performed via a hold-out method. Each dataset was divided based on each patient. The accuracy and AUC were calculated to evaluate model classification accuracy. The GradCAM was used to evaluate the explainability. Our EfficientNetB0-based classifiers yielded an Acc of 0.76 (left, MLO), 0.79 (right, MLO), 0.75 (left, CC), and 0.77 (right, CC) and an AUC of 0.85 (left, MLO), 0.88 (right, MLO), 0.83 (left, CC), and 0.86 (right, CC). The AUC of all models was >0.8, and the visualization map by the GradCAM determined the calcifications globally. These results show the possibility of an explainability in the classification problem to differentiate the presence and absence of breast calcifications.