Graph regularized least squares regression for automated breast ultrasound imaging

Breast cancer, recognized as one of the most pervasive malignancies affecting females, manifests a perpetual escalation in its worldwide morbidity. Timely screening offers patients interventions that are not only more expeditious but also more efficacious. ABUS, as an automated breast ultrasound imaging technology, has the advantages of high resolution, reproducible results, and strong objectivity. However, manual interpretation of ABUS images can be highly tedious and laborious when handling huge amounts of data from images, may leading to the risk of missed diagnosis. In practical clinical diagnosis, the incorporation of Computer-Aided Diagnosis (CAD) systems emerges as a more favorable option. In consideration of the holistic factors encompassing model construction efficiency, training cost and overall performance, we provide a unique subspace clustering approach known as Graph regularized Least Squares Regression (GLSR) in this paper. Firstly, we encode the data's intrinsic geometry into the framework of least squares regression to respect local and global correlation structures at the same time. Secondly, we construct a nearest neighbor graph to model the smooth structure of the representation, which may lead to a more accurate representation. Finally, a nonnegative constraint is introduced into our framework so that the representation obtained by GLSR has clearer physical meaning. We provide an optimization approach for solving such a framework based on iterative updates of four blocks. Extensive experimental results indicate the effectiveness of the proposed method compared with state-of-the-art approaches on real-world breast cancer images and other biological data. • By exploiting the representation to describe the nearest neighbor relationships in the data space and modeling the global structure of the data using the least squares technique, the proposed GLSR can capture both the global and local structures of the data. • The proposed GLSR is especially well-suited for data drawn from a union of multiple linear subspaces embedded in a high-dimensional space. • The proposed GLSR can not only be utilized in the diagnosis of breast cancer but also extended to more medical diagnosis scenarios.