Detecting and accurately locating kidney stones, which are common urological conditions, can be challenging when using imaging examinations. Therefore, the primary objective of this research is to develop an ensemble model that integrates segmentation and registration techniques. This model aims to visualize the inner structure of the kidney and accurately identify any underlying kidney stones. To achieve this, three separate datasets, namely non-contrast computed tomography (CT) scans, corticomedullary CT scans, and CT excretory scans, are annotated to enhance the three-dimensional (3D) reconstruction of the kidney’s complex anatomy. Initially, the research focuses on utilizing segmentation models to identify and annotate specific classes within the annotated datasets. Subsequently, a registration algorithm is employed to align and combine the segmented results, resulting in a comprehensive 3D representation of the kidney’s anatomical structure. Three cutting-edge segmentation algorithms are employed and evaluated during the segmentation phase, with the most accurate segments being selected for the subsequent registration process. Ultimately, the registration process successfully aligns the kidneys across all three phases and combines the segmented labels, producing a detailed 3D visualization of the complete kidney structure. For kidney segmentation, Swin UNETR exhibited the highest Dice score of 95.21%; for stone segmentation, ResU-Net achieved the highest Dice score of 87.69%. Regarding Artery, Cortex, and Medulla segmentation, ResU-Net and 3D U-Net show comparable performance with similar Dice scores. Considering the Collecting System and Parenchyma, ResU-Net and 3D U-Net demonstrate similar performance in Dice scores. In conclusion, the proposed ensemble model shows potential in accurately visualizing the internal structure of the kidney and precisely localizing kidney stones. This advancement improves the diagnosis process and preoperative planning in percutaneous nephrolithotomy.

使用影像检查时，检测并准确定位肾结石（常见的泌尿系统疾病）可能具有挑战性。因此，本研究的主要目标是开发一种集成分割和配准技术的集成模型。该模型旨在可视化肾脏的内部结构并准确识别任何潜在的肾结石。为了实现这一目标，对三个独立的数据集（即非对比计算机断层扫描 (CT) 扫描、皮质髓质 CT 扫描和 CT 排泄物扫描）进行注释，以增强肾脏复杂解剖结构的三维 (3D) 重建。最初，研究重点是利用分割模型来识别和注释注释数据集中的特定类别。随后，采用配准算法来对齐和组合分割结果，从而获得肾脏解剖结构的全面 3D 表示。在分割阶段采用和评估三种尖端分割算法，并选择最准确的分割用于后续配准过程。最终，配准过程成功地对齐所有三个阶段的肾脏，并结合分段标签，生成完整肾脏结构的详细 3D 可视化。对于肾脏分割，Swin UNETR 表现出最高的 Dice 分数，达到 95.21%；对于石头分割，ResU-Net 取得了最高的 Dice 分数，达到 87.69%。关于动脉、皮质和髓质分割，ResU-Net 和 3D U-Net 显示出具有相似 Dice 分数的可比性能。考虑到收集系统和实质，ResU-Net 和 3D U-Net 在 Dice 分数中表现出相似的性能。总之，所提出的整体模型显示了准确可视化肾脏内部结构和精确定位肾结石的潜力。这一进步改进了经皮肾镜取石术的诊断过程和术前计划。