Determining the Differentiation of Benign and Malignant NME Lesions in Contrast-Enhanced Spectral Mammography Images Based on Convolutional Neural Networks

PurposeContrast-enhanced spectral mammography (CESM) shows the contents of breast tissue with appropriate sensitivity, which include benign and malignant symptoms and lesions. In Mass, the lesion is certain, but in non-mass enhancement (NME) symptoms, there is no uniform pattern and structure to distinguish between benign and malignant cases. Thus, the goal is to use the BI-RADS standard, which has been evaluated using radiographic data and deep learning models to determine the difference between benign and malignant NME lesions.MethodA total of 184 lesions with NME have been investigated with a distribution of 73 benign and 111 malignant. Determining the structure of NME lesions according to BI-RADS standards was performed by radiologists. Image processing techniques were applied to improve the data quality. The suspicious area was manually separated from other breast tissue in MATLAB software. Finally, processing and extracting data features to improve the performance, the transfer learning methodology was used, which was performed by three pre-trained models Resnet-50, Resnet-18, and Densenet-201; the process of data training and classification was carried out by applying the K-Fold10 technique.ResultsIn the obtained results, segmental, regional, and linear morphological distribution and clumped and heterogeneous patterns have a significant relationship with the degree of malignancy according to Fisher's exact test and odds ratio (OR). In the evaluation of the proposed method from the three proposed models of the convolutional neural network with the transfer learning approach that was used, the proposed Densenet-201 model was able to perform well with sensitivity, specificity, and accuracy values of 100%, 91.25%, and 96%, respectively.ConclusionEarly diagnosis and prognosis of NME can play a significant role in the treatment and survival of affected people; the combination of clinical information and deep learning models can be very efficient and effective.