Diagnosis of architectural distortion on digital breast tomosynthesis using radiomics and deep learning

• Published in: Frontiers in oncology 2022-12, Vol.12, p.991892-991892
• Main Authors: Chen, Xiao, Zhang, Yang, Zhou, Jiahuan, Wang, Xiao, Liu, Xinmiao, Nie, Ke, Lin, Xiaomin, He, Wenwen, Su, Min-Ying, Cao, Guoquan, Wang, Meihao
• Format: Article
• Language: English
• Subjects: architectural distortion; breast cancer diagnosis; deep learning; digital breast tomosynthesis; Oncology; radiomics
• ISSN: 2234-943X; 2234-943X
• DOI: 10.3389/fonc.2022.991892

To implement two Artificial Intelligence (AI) methods, radiomics and deep learning, to build diagnostic models for patients presenting with architectural distortion on Digital Breast Tomosynthesis (DBT) images. A total of 298 patients were identified from a retrospective review, and all of them had confirmed pathological diagnoses, 175 malignant and 123 benign. The BI-RADS scores of DBT were obtained from the radiology reports, classified into 2, 3, 4A, 4B, 4C, and 5. The architectural distortion areas on craniocaudal (CC) and mediolateral oblique (MLO) views were manually outlined as the region of interest (ROI) for the radiomics analysis. Features were extracted using PyRadiomics, and then the support vector machine (SVM) was applied to select important features and build the classification model. Deep learning was performed using the ResNet50 algorithm, with the binary output of malignancy and benignity. The Gradient-weighted Class Activation Mapping (Grad-CAM) method was utilized to localize the suspicious areas. The predicted malignancy probability was used to construct the ROC curves, compared by the DeLong test. The binary diagnosis was made using the threshold of ≥ 0.5 as malignant. The majority of malignant lesions had BI-RADS scores of 4B, 4C, and 5 (148/175 = 84.6%). In the benign group, a substantial number of patients also had high BI-RADS ≥ 4B (56/123 = 45.5%), and the majority had BI-RADS ≥ 4A (102/123 = 82.9%). The radiomics model built using the combined CC+MLO features yielded an area under curve (AUC) of 0.82, the sensitivity of 0.78, specificity of 0.68, and accuracy of 0.74. If only features from CC were used, the AUC was 0.77, and if only features from MLO were used, the AUC was 0.72. The deep-learning model yielded an AUC of 0.61, significantly lower than all radiomics models (p