| 孙兆男,何江凯,王可欣,黄文鹏,吴鹏升,张晓东,王霄英.3D V-Net深度学习模型用于自动分割T2WI及表观弥散系数图所示前列腺[J].中国医学影像技术,2024,40(9):1426~1431 |
| 3D V-Net深度学习模型用于自动分割T2WI及表观弥散系数图所示前列腺 |
| 3D V-Net deep learning model for automatic segmentation of prostate on T2WI and apparent diffusion coefficient maps |
| 投稿时间：2023-10-16 修订日期：2024-06-10 |
| DOI：10.13929/j.issn.1003-3289.2024.09.031 |
| 中文关键词: 前列腺 磁共振成像 人工智能 自动分割 |
| 英文关键词:prostate magnetic resonance imaging artificial intelligence automated segmentation |
| 基金项目: |
|
| 摘要点击次数: 518 |
| 全文下载次数: 228 |
| 中文摘要: |
| 目的 基于多中心数据开发3D V-Net深度学习模型,观察其自动分割T2WI及表观弥散系数(ADC)图所示前列腺的价值。方法 回顾性收集来自3个医疗中心、于穿刺活检前1个月内接受2 894次多参数MRI的2 673例疑诊前列腺癌患者;纳入其中5 974组轴位图像,包括3 654组 T2WI及2 320组ADC图。于轴位T2WI和ADC图中逐层手动标注前列腺轮廓,测量前列腺左右径、前后径、上下径及体积作为参考标准。按8 ∶ 1 ∶ 1比例将全部图像分为训练集(n=4 780,含2 907组T2WI及1 873组ADC图)、验证集(n=601,含384组T2WI及217组ADC图)与测试集(n=593,含363组T2WI及230组ADC图),行预处理及扩增后,采用3D V-Net基于训练集和验证集构建及训练分割模型,于测试集以戴斯相似系数(DSC)、杰卡德系数(JACARD)及体积相似度(VS)评估模型分割效能;比较模型所测前列腺各参数与参考标准的差异并分析其相关性。结果 相比对应ADC图,模型自动分割测试集T2WI中前列腺的DSC、JACARD、VS均较高(P均＜0.001)。模型所测T2WI及ADC图中的前列腺左右径、前后径、上下径均大于参考标准(P均＜0.001)而体积差异均无统计学意义(P均＞0.05);模型基于T2WI及ADC图所测前列腺各参数均与参考标准呈正相关(rs=0.794~0.985)。结论 3D V-Net深度学习模型自动分割T2WI及ADC图所示前列腺的准确性较高,且基于T2WI的效能优于ADC图。 |
| 英文摘要: |
| Objective To develop a 3D V-Net deep learning segmentation model based on multi-center data, and to evaluate its value for automatic segmentation of prostate on T2WI and apparent diffusion coefficient (ADC) maps. Methods Totally 2 894 sets of multi-parametric MRI data of 2 673 patients with clinically suspected prostate cancer from 3 medical centers within 1 month before biopsy were retrospectively collected. Finally 5 974 sets axial images were enrolled, including 3 654 sets of T2WI and 2 320 sets of ADC maps. Prostate contours were manually annotated layer by layer on axial T2WI and ADC maps, and the left-to-right, anterior-to-posterior, superior-to-inferior diameters and volume of prostate were measured and taken as reference standards. The images were divided into training set (n=4 780, including 2 907 sets of T2WI and 1 873 sets of ADC map), verification set (n=601, including 384 sets of T2WI and 217 sets of ADC map) and test set (n=593, including 363 sets of T2WI and 230 sets of ADC map) at the ratio of 8 ∶ 1 ∶ 1. After preprocessing and augmentation, 3D V-Net was used to construct and train the segmentation model based on training and verification sets, and the segmentation performance of the model was evaluated in test set using Dice similarity coefficient (DSC), Jaccard coefficient (JACARD) and volume similarity (VS), respectively. The parameters measured with the model were compared with the reference standards, and the correlations were explored. Results Compared with the corresponding ADC maps, DSC, JACARD and VS of the model for automatic segmentation of prostate on T2WI in test set were all higher (all P＜0.001). The left-to-right, anterior-to-posterior and superior-to-inferior diameters of prostate measured with the model on both T2WI and ADC maps were all larger than the reference standards (all P＜0.001), while no significant difference of the volume was found (both P＞0.05). All parameters measured with the model on T2WI and ADC maps were positively correlated with reference standards (rs=0.794—0.985). Conclusion 3D V-Net deep learning model could automatically segment prostate on T2WI and ADC maps with high accuracy, and its efficiency based on T2WI was better than that based on ADC maps. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
