摘要
目的使用深度学习模型实现影像诊断为"胰腺未见异常"的成人群体CT图像中胰腺自动分割及体积、径线和平均CT值自动测量。方法搜集2019年1月至10月在本院行腹盆部CT平扫及增强扫描并诊断为"胰腺未见异常"的患者共1195例(8301个不同期相、不同层厚的扫描序列,每个序列为一个图像数据),根据标准共纳入了5389个图像数据。将图像数据分为两部分:第一部分521个图像数据由专家标注后用于3D U-Net分割模型训练,将其随机分为训练集(413个数据)、调优集(56个数据)、测试集(52个数据),模型训练的评价指标为测试集的DICE系数;第二部分4868个图像数据用于模型外部验证,使用第一部分训练好的模型预测胰腺区域,由两位影像专家检查模型预测结果,挑选出分割效果满意的数据共2003个。对上述521及2003个数据中胰腺标签进行数据处理,以最小体积包围盒算法测量胰腺的三维径线,以标签区域的像素总体积计算胰腺体积,并输出标签区域胰腺的平均CT值。统计各期相、不同层厚的图像数据中胰腺体积、三维径线及平均CT值的95%参考值范围,分析每10岁年龄组胰腺体积分布的正常范围。结果胰腺分割模型测试集中DICE值为0.91。胰腺三维径线平均值分布范围分别为41.97~43.51 mm, 65.84~71.09 mm和137.77~142.59 mm^(3)胰腺体积平均值的分布范围为65213.35~69864.79 mm^(3),在18~38岁逐渐平均体积增大,38岁平均值(77963.15±15125.93) mm^(3),之后随年龄增大体积逐渐减小,在78~88岁时平均体积最低,为(51349.88±18998.81) mm^(3)。胰腺CT值则随年龄增大而减小,18岁~88岁平均CT值从(39.22±9.57) HU降低至(23.95±6.87) HU。结论基于深度学习的人工智能(AI)分割工具可以有效分割CT图像上的胰腺,并准确测量其径线、体积及CT值。
Objective To use a deep learning model to achieve automatic segmentation of pancreas and measurement of volume, diameter, and CT value on CT images in adults diagnosed with“normal pancreas” according to CT images. Methods A total of 1195 CT examinations(including 8301 different phases or image thickness in total) were collected from January 2019 to October 2019.5389 image data met the entry criteria and were included in this research.The process was divided into two parts.First, 521 image data were annotated by two experts and used for training of a 3 D U-Net segmentation model.They were randomly divided into training sets(n=413),validation set(n=52) and test set(n=56).DICE coefficient of the test dataset was used to evaluate the efficacy of the model. Secondly, the other 4868 image data were used for external verification of the model and two experts checked the segmentation results and selected 2003 data with satisfactory segmentation effects. The data processing was performed on the above 521 and 2003 data. Using the minimum volume bounding box algorithm to measure the diameters of the pancreas to calculate the pancreatic volume and the average CT value based on the total pixel volume of the pancreas. The 95% reference value range of pancreatic volume, diameters and average CT value of each phase and different slice thickness were calculated, and the volumes were analyzed in every 10-year-old age group. Results The DICE of the pancreas segmentation model in the test dataset was 0.91. The diameters of the pancreas were 41.97-43.51 mm, 65.84-71.09 mm and 137.77-142.59 mm, respectively. The average pancreatic volume was 65213.35-69864.79 mm^(3),which increased to the maximum 77963.15 mm^(3)at the age of 18-38 groups, and decreased to the lowest value at the age of 78-88 group to 51349.88 mm^(3).The CT value of the pancreas decreased with age from the maximum value of 39.22 HU to 23.95 HU. Conclusion The AI segmentation tool based on deep learning can segment the pancreas on CT images and measure its diameter,
作者
蔡金秀
崔应谱
孙兆男
张耀峰
张大斗
张晓东
王霄英
CAI Jinxiu;CUI Yingpu;SUN Zhaonan(Department of Radiology,Peking University First Hospital,Beijing 100034,P.R.China)
出处
《临床放射学杂志》
北大核心
2022年第1期86-93,共8页
Journal of Clinical Radiology
