A multi-center study of ultrasound images using a fully automated segmentation architecture

Accurate organ segmentation in ultrasound (US) images remains challenging because such images have inhomogeneous intensity distributions in their regions of interest (ROIs) and speckle and imaging artifacts. We address this problem by developing a coarse-to-refinement architecture for the segmentation of multiple organs (i.e., the prostate and kidney) in US image datasets from multiple centers. Our proposed architecture has the following four advantages: (1) it inherits the ability of the deep learning models to locate an ROI automatically while also using a principal curve approach to automatically fit a dataset center; (2) it takes advantage of a principal curve-based enhanced polygon searching method, which inherits the principal curve's characteristic to automatically approach the center of the dataset; (3) it incorporates quantum characteristics into a storage-based evolution network together to improve the global search performance of our method, which includes several improvements, such as a new quantum mutation module, a cuckoo search method, and global optimum schemes; (4) it incorporates a suitable mathematical model to smooth the contour of ROIs, which is explained by the parameters of a neural network model. Application of our method to US image datasets of multiple organs and from multiple centers demonstrates that it achieves satisfactory segmentation performance.