Diagnosis of Benign and Malignant Breast Lesions on DCE-MRI by Using Radiomics and Deep Learning With Consideration of Peritumor Tissue.

摘要: Background Computer-aided methods have been widely applied to diagnose lesions detected on breast MRI, but fully-automatic diagnosis using deep learning is rarely reported. Purpose To evaluate the diagnostic accuracy of mass region interest (ROI)-based, radiomics and deep-learning methods, by taking peritumor tissues into consideration. Study type Retrospective. Population In all, 133 patients with histologically confirmed 91 malignant 62 benign for training (74 48 26 testing). Field strength/sequence 3T, volume imaging assessment (VIBRANT) dynamic contrast-enhanced (DCE) sequence. Assessment 3D tumor segmentation was done automatically fuzzy-C-means algorithm connected-component labeling. A total 99 texture histogram parameters were calculated each case, 15 selected random forest build a model. Deep implemented ResNet50, evaluated 10-fold crossvalidation. The alone, smallest bounding box, 1.2, 1.5, 2.0 times enlarged boxes used as inputs. Statistical tests malignancy probability model, threshold 0.5 make diagnosis. Results dataset, 76% three ROI-based parameters, 84% 86% ROI + per-slice basis, area under receiver operating characteristic (ROC) comparable 1.2 box (AUC = 0.97-0.99), which significantly higher than 1.5 0.86 0.71, respectively). For per-lesion diagnosis, highest 91% achieved when that decreased alone further 73% 69% box. independent testing also 89%. Data conclusion ResNet50 high accuracy. Using containing proximal tissue input had compared or larger boxes. Level evidence 3 Technical Efficacy: Stage 2.