Assessing the Efficiency of Deep Learning Methods in Estimating the Malignancy of Bi-Rads 4 Breast Lesions Using Contrast-enhanced Spectral Mammography Images

According to the Breast Imaging-Reporting and Data System (BI-RADS), category 4 breast lesions have a 2-95% probability of malignancy. Such estimation can cause challenges in planning for the treatment of women with breast cancer. Contrast-enhanced spectral mammography (CESM) is one of the best imaging modalities in breast cancer detection. In this study, we aim to assess the efficiency of deep learning methods in determining the malignancy degree of BI-RADS 4 breast lesions using CESM images.

In this study, 1408 CESM images of BI-RADS 4 breast lesions were used. The image pre-processing step was first done to remove noises and improve image quality. Then, segmentation was done for the region of interest extraction. Feature extraction was done using three different conventional classifiers. Finally, the classification of images was done using deep learning methods.

Among the applied methods, the Densenet-201 network used for feature extraction and K-nearest neighbor (KNN) used for Classification showed the best results with accuracy, sensitivity, specificity, and area under the curve of 98.57%, 99.20%, 97.50% and 0.987 respectively.

The proposed method (Densenet-201 and KNN) using CESM images is effective in estimating the malignancy of BI-RADS 4 breast lesions and thus in timely treatment of breast cancer.