Usefulness of Deep Learning Models of DWI in Direct Differentiating Malignant and Benign Breast Tumors Without Lesion Segmentation

348 women suspected of breast tumors were enrolled, and 206 breast lesions (139 malignant, 67 benign) were further analyzed. Breast 5b-value-DWI was performed, and a deep learning model dedicated to this breast DWI dataset was established. Several comparative experiments were performed; comparison of data augmentations, small network VS. large network, 2DCNN VS. 3DCNN, and analysis in 5b-DWI VS. ADC maps. Augmentations using elastic deformation, affine transform, and Gaussian noise improved diagnostic performance up to AUC=0.90. The use of small CNN and without ADC map also showed higher diagnostic performance (AUC=0.88-0.90), showing AI potential to improve breast DWI diagnostic performance.