深度学习计算T2加权像磁共振前列腺体积:与椭球公式比较

Deep Learning for Calculation of the Prostate Volume in T2-Weighted MR Images:Comparison with Pro-late Ellipsoid Formula

目的研究和实现利用深度学习计算T2加权像MR的前列腺体积,并与椭球公式计算的前列腺体积进行比较。资料与方法回顾性收集2019年10月—2022年2月武汉大学人民医院经病理确诊的180例前列腺增生和251例前列腺癌患者的T2加权像MR图像及诊断报告,根据诊断报告使用椭球公式计算每例患者的前列腺体积,使用U-Net模型的变体对所收集MR图像上的前列腺进行分割,利用公式前列腺体积=对(前列腺像素数目×每个像素的大小×层厚)进行求和,获得深度学习计算的前列腺体积。比较深度学习和椭球公式计算的前列腺体积差异和一致性。结果Bland-Altman分析显示,在前列腺增生和前列腺癌患者中,深度学习和椭球公式计算的前列腺体积具有较高的一致性,仅5%和6.37%的数据位于95%置信区间外。前列腺增生组用两种方法计算的前列腺体积的一致性高于前列腺癌组(ICC=0.803、0.686)。两种方法计算的前列腺体积在两组间差异有统计学意义(Z=-10.742、-12.706,P<0.05),深度学习计算的前列腺体积更大。结论深度学习在计算前列腺体积方面与椭球公式保持一致,利用深度学习计算MR前列腺体积具有广阔的前景,但还需进一步改进。

Purpose To investigate and implement the utilization of deep learning for calculating the prostate volume in T2-weighted MR images,and to compare it with the prostate volume calculated using the prolate ellipsoid formula.Materials and Methods T2-weighted MR images and diagnostic reports of 180 patients pathologically confirmed benign prostatic hyperplasia and 251 patients with pathologically confirmed prostate cancer in Renmin Hospital of Wuhan University were collected from October 2019 to February 2022.The prostate volume was calculated for each patient based on the diagnostic report using the prolate ellipsoid formula.The prostate was segmented using a U-Net-based deep learning model.The formula,prostate volume=sum(number of pixels in the prostate×size of each pixel×thickness),was used to obtain the prostate volume calculated using deep learning.The difference and consistency between the prostate volume calculated using deep learning and the prolate ellipsoid formula were compared.Results Bland-Alteman analysis revealed that the prostate volume calculated using the two methods in benign prostatic hyperplasia and prostate cancer showed high consistency,with only 5%and 6.37%of the data,respectively,falling outside the 95%confidence interval.Prostate volume consistency was higher in benign prostatic hyperplasia than in prostate cancer(ICC=0.803,0.686).There was a significant difference in prostate volume calculated by the two methods in both groups(Z=-10.742,-12.706,P<0.05),with a larger prostate volume calculated using deep learning.Conclusion Deep learning remains consistent with the prolate ellipsoid formula in calculating prostate volume.Therefore,utilizing deep learning for calculating MR prostate volume holds vast prospects,but further improvement is needed.