<p>The kidney contour is a critical reference for assessing whether a renal tumor has penetrated the renal capsule and invaded adjacent tissues. Accurate segmentation of the kidney contour is essential for estimating renal volume and constructing patient-specific anatomical models, which are vital for pre-operative planning and image-guided biopsy. Manual delineation of the kidney contour is time-consuming and prone to inter- and intra-observer variations. Addressing the development of novel methods for precise kidney boundary delineation in ultrasonic data is challenging owing to the absence or indistinctness of these boundaries. In this work, a novel segmentation method named a Caputo principal-curve-guided segmentation network (CPSN) integrated principal-curve (PC)-based vertex decision into a Caputo multiple-layer learning network was developed to boost the precision of ultrasound (US) kidney segmentation. First, an initial deep network was designed to extract the rough contour information. Second, the PC-based vertex decision block was adopted to determine the distribution order of vertices, which was subsequently used as the input of the Caputo multiple-layer training network. Third, the Caputo training network was further trained to decrease the global model error and improve alignment between predicted results and ground truth labels. Several experiments were conducted to validate the effectiveness and robustness of our method on multi-institute US kidney databases, which were proven to achieve superior performance compared to other state-of-the-art (SOTA) techniques, with Dice index (DI), Jaccard index (JI), and accuracy (ACC) values of 94.6 ± 3.2%, 93.4 ± 3.7%, and 94.1 ± 3.47%, respectively.</p>