Institute of Fundamental Technological Research

In this work we investigate the usefulness and robustness of transfer learning with deep convolutional neural networks (CNNs) for breast lesion classification in ultrasound (US). Deep learning models can be vulnerable to adversarial examples, engineered input image pixel intensities perturbations that force models to make classification errors. In US imaging, distribution of US image pixel intensities relies on applied US image reconstruction algorithm. We explore the possibility of fooling deep learning models for breast mass classification by modifying US image reconstruction method. Raw radio-frequency US signals acquired from malignant and benign breast masses were used to reconstruct US images, and develop classifiers using transfer learning with the VGG19, InceptionV3 and InceptionResNetV2 CNNs. The areas under the receiver operating characteristic curve (AUCs) obtained for each deep learning model developed and evaluated using US images reconstructed in the same way were equal to approximately 0.85, and there were no associated differences in AUC values between the models (DeLong test p-values > 0.15). However, due to small modifications of the US image reconstruction method the AUC values for the models utilizing the VGG19, InceptionV3 and InceptionResNetV2 CNNs significantly decreased to 0.592, 0.584 and 0.687, respectively. Our study shows that the modification of US image reconstruction algorithm can have significant negative impact on classification performance of deep models. Taking into account medical image reconstruction algorithms may help develop more robust deep learning computer aided diagnosis systems.

Adversarial attacks, Breast lesion classification, Computer aided diagnosis, Deep learning, Robustness, Ultrasound imaging, Transfer learning