Discrete Wavelet Coefficient-based Embeddable Branch for Ultrasound Breast Masses Classification

Abstract

The progress of computer-aid-diagnosis system for ultrasound breast lesions reaches tremendous success in the past few years. However, conventional deep learning-based strategies in recent developments still have challenges particularly in characterizing tumor domain in ultrasound images due to the heterogeneous and complex variations of lesions along with similar intensity exhibited in target object. To address this, this work proposes a discrete wavelet coefficient-based embeddable branch that allows to additionally propagate geometrical features of tumors in an end-to-end trainable fashion. To be elaborate, such branch priorly enforce the wavelet pooling operation to select a certain coefficient to further collect gradient information of target domain. Further, the current work also investigates two different preprocessing strategies in which the internal and external gradients of lesion areas can be emphasized within the transformation. Thus, we examine the effects of the proposed method based on different preprocessing scenarios. To verify the usefulness, GradCam projection, and the cross-validation demonstrate the connection of the proposed branch encourages the importance of target features, thus boosting the overall discrimination between lesion groups. Lastly, the proposed branch can be easily incorporated with existing deep learning-based architectures.