| 李桂玉,马温惠,汪俊伶,马涛奇,王云雅,康飞,杨卫东,汪静.深度学习模型自动分割基于一体化68Ga-前列腺特异性膜抗原PET/MRI大视野T2WI中的前列腺[J].中国医学影像技术,2024,40(10):1588~1592 |
| 深度学习模型自动分割基于一体化68Ga-前列腺特异性膜抗原PET/MRI大视野T2WI中的前列腺 |
| Deep learning model for automatically segmenting prostate onlarge-field T2WI based on integrated 68Ga-prostatespecific membrane antigen PET/MRI |
| 投稿时间：2024-05-08 修订日期：2024-08-26 |
| DOI：10.13929/j.issn.1003-3289.2024.10.027 |
| 中文关键词: 前列腺肿瘤 深度学习 正电子发射断层显像 磁共振成像 |
| 英文关键词:prostatic neoplasms deep learning positron-emission tomography magnetic resonance imaging |
| 基金项目: |
|
| 摘要点击次数: 434 |
| 全文下载次数: 215 |
| 中文摘要: |
| 目的 观察深度学习模型自动分割基于一体化68Ga-前列腺特异性膜抗原(PSMA)PET/MRI的大视野T2WI中的前列腺的价值。方法 回顾性收集90例接受68Ga-PSMA PET/MRI的男性前列腺肿瘤患者,按4:1将其分为训练集(n=72)与验证集(n=18)。分别基于3D SegResNet和3D Unet深度学习神经网络建立模型,并以医师人工分割结果为参考标准,评估模型分割大视野T2WI所示前列腺外周带(PZ)及中央带(CZ)+移行带(TZ)的效能。结果 训练集和验证集中,3D SegResNet深度学习模型分割T2WI中的前列腺的戴斯相似系数(DSC)均大于3D Unet模型(P均＜0.05)、分割前列腺CZ+TZ的95%豪斯多夫距离(HD95)均小于3D Unet模型(P均＜0.05);且2个模型分割前列腺CZ+TZ的DSC和HD95均优于PZ(P均＜0.05)。结论 3D SegResNet深度学习模型可较好地自动分割基于一体化68Ga-PSMA PET/MRI大视野T2WI中的前列腺。 |
| 英文摘要: |
| Objective To observe the value of deep learning model for automatically segmenting prostate on large-field T2WI based on integrated 68Ga-prostate specific membrane antigen (PSMA) PET/MRI. Methods Ninety male patients with prostate tumors who underwent 68Ga-PSMA PET/MRI were retrospectively enrolled and divided into training set (n=72) and validation set (n=18) at the ratio of 4∶1. Models were established based on 3D SegResNet and 3D Unet deep learning neural networks, respectively. Taken physicians' manual segmentation results as reference standards, the performances of models for segmenting the peripheral zone (PZ) and central zone (CZ)+transition zone (TZ) of prostate on large-field T2WI were evaluated. Results In both training and validation sets, the Dice similarity coefficient (DSC) of 3D SegResNet deep learning model for segmenting prostate on T2WI were both higher than that of 3D Unet model (both P＜0.05), the 95% Hausdorff distance (HD95) of SegResNet deep learning model for segmenting prostate CZ+TZ was lower than that of 3D Unet model (both P＜0.05), while DSC and HD95 of these 2 models for segmenting prostate CZ+TZ were superior to PZ (all P＜0.05). Conclusion 3D SegResNet deep learning model could be used to automatically segment prostate on large-field T2WI based on integrated 68Ga-PSMA PET/MRI. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
