PURPOSE: 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. MATERIALS AND METHODS: 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. RESULTS: 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<0.01), which was presumably due to the use of the entire image as input. The Grad-CAM could localize the architectural distortion areas. CONCLUSION: The radiomics model can achieve a satisfactory diagnostic accuracy, and the high specificity in the benign group can be used to avoid unnecessary biopsies. Deep learning can be used to localize the architectural distortion areas, which may provide an automatic method for ROI delineation to facilitate the development of a fully-automatic computer-aided diagnosis system using combined AI strategies.

A total of 457 patients, including 241 HR+/HER2- patients, 134 HER2+ patients, and 82 TN patients, were studied. The percentage of TILs in the stroma adjacent to the tumor cells was assessed using a 10% cutoff. The low TIL percentages were 82% in the HR+ patients, 63% in the HER2+ patients, and 56% in the TN patients (p < 0.001). MRI features such as morphology as mass or non-mass enhancement (NME), shape, margin, internal enhancement, presence of peritumoral edema, and the DCE kinetic pattern were assessed. Tumor sizes were smaller in the HR+/HER2- group (p < 0.001); HER2+ was more likely to present as NME (p = 0.031); homogeneous enhancement was mostly seen in HR+ (p < 0.001); and the peritumoral edema was present in 45% HR+, 71% HER2+, and 80% TN (p < 0.001). In each subtype, the MR features between the high- vs. low-TIL groups were compared. In HR+/HER2-, peritumoral edema was more likely to be present in those with high TILs (70%) than in those with low TILs (40%, p < 0.001). In TN, those with high TILs were more likely to present a regular shape (33%) than those with low TILs (13%, p = 0.029) and more likely to present the circumscribed margin (19%) than those with low TILs (2%, p = 0.009).