Deep Learning-Based Breast Tumor Classification Using Shear-Wave Sonoelastography Image Features and Clinical...

It has been shown that shear-wave elastography (SWE) can help the radiologist with the early and efficient classification of breast masses. This study aimed to investigate the role of clinical variables and SWE image features in the classification of breast masses using several common Convolutional Neural Networks (CNNs).

834 SWE images were collected prospectively, and of these, 534 images were used to train popular CNNs, including ResNet101, VGG16, Exception, InceptionV3, and DenseNet169. Region of interest (ROI) on SWE images was obtained from a 2D mask manually cropped on B-mode images. Classification of images with and without ROI selection was performed once using image features and once using clinical variables along with image features, and the models' performance was evaluated.

The best performance was related to the DensNet169 and ResNet152 in the classification using both clinical variables and SWE image features such that the accuracy, AUC_Test, and AUC_Validation of DensNet169 were 94.01%, 0.86, and 0.97, respectively, and for ResNet152 were 91.62%, 0.83 and 0.98. No significant difference was observed in the performance of CNNs for classifying images with and without ROI selection. However, the AUCs of DensNet169, VGG16, and InceptionV3 for images with ROI selection were significantly improved using both clinical variables and image features (P ≤ 0.05).

We conclude that using clinical variables significantly improved the performance of most CNNs in classifying masses on images with ROI selection.