%0 Journal Article %T BREAST TUMOR CLASSIFICATION IN SHEAR WAVE ELASTOGRAPHY IMAGES USING CONVOLUTIONAL NEURAL NETWORK %J Iranian Congress of Radiology %I Iranian Society of Radiology %Z 25885545 %A Ahmadinejad, Nasrin %A Ghelich Oghli, Mostafa %A Moradi, Shakiba %A Aryan, Arvin %A Shiri, Isaac %D 2019 %\ 09/01/2019 %V 35 %N 4 %P 96-96 %! BREAST TUMOR CLASSIFICATION IN SHEAR WAVE ELASTOGRAPHY IMAGES USING CONVOLUTIONAL NEURAL NETWORK %R 10.22034/icrj.2019.100891 %X Abstract Background: Breast cancer is the most common type of cancer among women. About one in eight women are diagnosed with breast cancer during their lifetime. A malignant tissue is stiffer than normal and benign tissues. This stiffness could be evaluated by elastography. Patients and Methods: A comprehensive dataset of shear wave elastography (SWE) images of breast tissue using an Esaote MyLab™ 9 system was provided. 100 images were related to breasts with benign tumors and 100 images contain malignant tumors. The gold standard for evaluation of proposed algorithm was biopsy, which was performed on all of examining lesions. A convolutional neural network was applied to the dataset, to extract the visual features of the images. The architecture was based on Densenet architecture, which is modified for our purpose. We have used the network in both pre-trained and end-to-end training strategies and the results were compared. The network was pre-trained on the Imagenet dataset, due to the lack of sufficient dataset. On the other hand, with data augmentation the network underwent  a full training strategy. Finally, the classification layer, which decides about the benignity or malignancy of the lump, is a softmax layer.   Results: The results of the proposed methods are satisfying in both pre-trained and end-to-end training approaches. We have used various evaluation metrics contain precision, recall, F1-score, ROC curve, and training time for both strategies. The precision, recall, and F1-score were 0.87, 0.91, and 0.89 for the Densenet architecture trained from scratch and 0.93, 0.95,  and 0.94 for the transfer learning approach. The ROC curves were plotted for both approaches and the area under the curves (AUCs) were calculated. The transfer learning approach yielded a 0.97 of AUC, whereas this number was 0.91 for the fully- trained approach. At last, the training time of transfer learning approach was less than training from scratch, as it was anticipated.   Conclusion: The results show the superiority of the transfer learning approach in tumor classification. Higher statistical metrics with lower training time makes this approach more compatible with SWE images. %U