文献速递腹部泌尿生殖技术与安全CT

基于深度学习的CT图像肾脏肿瘤自动分割:一种多阶段方法

Automated Kidney Tumor Segmentation in CT Images Using Deep Learning: A Multi-Stage Approach.
PubMed 文献 · Acad Radiol · 2025-09-03

本研究开发了基于DeepMedic三维卷积神经网络的全自动分割流程,通过多尺度特征提取对肾脏和肾肿瘤进行分割。在382例经经验丰富医师手动标注的增强CT扫描上采用5折交叉验证训练和评估,肾脏分割平均Dice系数为93.82±1.38%,肾肿瘤分割Dice系数为88.19±1.24%。作者认为该框架能稳健、准确地分割CT图像上的肾脏和肾肿瘤,有望用于实时应用,提升肾肿瘤学的诊断效率和治疗计划。

要点

  • 肾脏分割平均Dice系数93.82±1.38%,精确率94.86±1.59%,召回率93.66±1.77%。
  • 肾肿瘤分割Dice系数88.19±1.24%,精确率90.36±1.90%,召回率88.23±2.02%。
  • 与真实标注的视觉比较证实了预测结果的临床相关性和准确性。

核心结论

  • 基于DeepMedic的框架在CT图像上对肾脏和肾肿瘤表现出稳健、准确的分割性能,具有实时应用潜力,可能提升肾肿瘤学的诊断效率和治疗计划。

研究设计与样本

  • 研究开发了基于DeepMedic三维卷积神经网络的全自动分割流程,通过多尺度特征提取分割肾脏和肾肿瘤;使用382例经经验丰富医师手动标注的增强CT扫描,采用5折交叉验证进行训练和评估;图像预处理包括Hounsfield单位转换、窗化、三维重建和体素重采样,并采用后处理优化输出掩膜以提高模型泛化能力。

诊断标准与关键数值

  • 肾脏分割:平均Dice系数93.82±1.38%,精确率94.86±1.59%,召回率93.66±1.77%。
  • 肾肿瘤分割:Dice系数88.19±1.24%,精确率90.36±1.90%,召回率88.23±2.02%。

临床应用与阅片要点

  • 作者认为该模型具有实时应用潜力,可能提升肾肿瘤学的诊断效率和治疗计划。

局限性

  • 原文未报告局限性

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

英文原文(摘要)· 点击展开

RATIONALE AND OBJECTIVES: Computed tomography (CT) remains the primary modality for assessing renal tumors; however, tumor identification and segmentation rely heavily on manual interpretation by clinicians, which is time-consuming and subject to inter-observer variability. The heterogeneity of tumor appearance and indistinct margins further complicate accurate delineation, impacting histopathological classification, treatment planning, and prognostic assessment. There is a pressing clinical need for an automated segmentation tool to enhance diagnostic workflows and support clinical decision-making with results that are reliable, accurate, and reproducible.

MATERIALS AND METHODS: This study developed a fully automated pipeline based on the DeepMedic 3D convolutional neural network for the segmentation of kidneys and renal tumors through multi-scale feature extraction. The model was trained and evaluated using 5-fold cross-validation on a dataset of 382 contrast-enhanced CT scans manually annotated by experienced physicians. Image preprocessing included Hounsfield unit conversion, windowing, 3D reconstruction, and voxel resampling. Post-processing was also employed to refine output masks and improve model generalizability.

RESULTS: The proposed model achieved high performance in kidney segmentation, with an average Dice coefficient of 93.82 ± 1.38%, precision of 94.86 ± 1.59%, and recall of 93.66 ± 1.77%. In renal tumor segmentation, the model attained a Dice coefficient of 88.19 ± 1.24%, precision of 90.36 ± 1.90%, and recall of 88.23 ± 2.02%. Visual comparisons with ground truth annotations confirmed the clinical relevance and accuracy of the predictions.

CONCLUSION: The proposed DeepMedic-based framework demonstrates robust, accurate segmentation of kidneys and renal tumors on CT images. With its potential for real-time application, this model could enhance diagnostic efficiency and treatment planning in renal oncology.

来源:Acad Radiol,查看原文。版权归原作者所有。

引用

  1. Kan HC, Fan GM, Wei MH, et al. Automated Kidney Tumor Segmentation in CT Images Using Deep Learning: A Multi-Stage Approach. Acad Radiol. 2025. PMID: 40908232. doi:10.1016/j.acra.2025.08.020

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。