经腹超声图像中前列腺体积的自动估计
本研究基于100例患者的经腹超声视频,构建深度学习框架自动分割前列腺并估算前列腺体积,与专家标注的直径计算作为参考标准比较。轴位模型分割表现优于矢状位模型,Dice分别为0.76±0.16与0.68±0.21,框架估算体积的平均体积误差为-2.1 mL(95%一致性界限-16.9至21.1 mL),相对误差小于25%。
要点
- 轴位模型分割Dice 0.76±0.16,优于矢状位模型的0.68±0.21。
- 框架估算前列腺体积的平均体积误差为-2.1 mL,95%一致性界限-16.9至21.1 mL。
- 体积估算的相对误差小于25%,提示深度学习可能用于无创前列腺体积估算。
核心结论
- 深度学习在经腹超声上自动估算前列腺体积具有潜力,可能支持改进前列腺癌风险分层。
研究设计与样本
- 研究纳入100例患者的经腹超声视频(中位年龄67岁,95百分位范围55-81.2岁),以专家勾画的前列腺边界和直径计算为参考标准,构建深度学习框架进行轴位和矢状位前列腺分割、自动直径估计和体积计算。
诊断标准与关键数值
- 轴位模型分割Dice为0.76±0.16,矢状位模型为0.68±0.21。
- 轴位模型Dice-MidPlane为0.91±0.06,矢状位模型为0.83±0.10。
- 轴位模型Hausdorff距离为6.21±4.33 mm,矢状位模型为7.93±4.27 mm。
- 框架估算前列腺体积的平均体积误差为-2.1 mL(95%一致性界限-16.9至21.1 mL)。
- 体积估算的相对误差小于25%。
临床应用与阅片要点
- 轴位模型的分割表现优于矢状位模型,阅片时可优先参考轴位分割结果。
- 框架估算体积的平均误差为-2.1 mL,报告时应了解其95%一致性界限为-16.9至21.1 mL。
局限性
- 经腹超声图像质量较低且依赖操作者,原文指出这是其相对于经直肠超声的局限。
- 原文结论称这些发现凸显深度学习用于准确无创前列腺体积估算的潜力,未报告其他局限性。
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
INTRODUCTION: Prostate cancer is a major health concern requiring accurate and accessible methods for early detection and risk stratification. Prostate volume (PV) is a key parameter in multivariate risk assessment, traditionally measured using transrectal ultrasound (TRUS). While TRUS provides precise measurements, its invasive nature affects patient comfort. Transabdominal ultrasound (TAUS) offers a non-invasive alternative but is limited by lower image quality and operator dependence. This study presents a deep-learning-based framework for automatic PV estimation using TAUS, aiming to improve non-invasive prostate cancer risk stratification.
METHODS: A dataset of TAUS videos from 100 patients (median age 67, 95-percentile range 55-81.2) was curated, with expert-delineated prostate boundaries and diameter calculations as ground truth. The framework integrates deep-learning models for prostate segmentation in both axial and sagittal planes, automatic diameter estimation, and PV calculation. Segmentation performance was evaluated using Dice correlation coefficient (%) and Hausdorff distance (mm), while volume estimation accuracy was assessed through volumetric error (mL).
RESULTS: The axial model outperformed the sagittal model, achieving a Dice score of 0.76 ± 0.16 versus 0.68 ± 0.21, a Dice-MidPlane of 0.91 ± 0.06 versus 0.83. ± 0.10, and a Hausdorff distance of 6.21 ± 4.33 mm versus 7.93 ± 4.27 mm. The framework estimated PV with a mean volumetric error of -2.1 mL (95 % limits of agreement: -16.9 to 21.1 mL), resulting in a relative error smaller than 25 %.
CONCLUSION: These findings highlight the potential of deep learning for accurate, non-invasive PV estimation, supporting improved prostate cancer risk assessment.
引用
- Natali T, Kurucz LM, Fusaglia M, et al. Automatic prostate volume estimation in transabdominal ultrasound images. Eur J Radiol. 2025. PMID: 40614658. doi:10.1016/j.ejrad.2025.112274
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。