文献速递：肿瘤分割----基于卷积神经网络的系统，用于前列腺癌[68Ga]Ga-PSMA PET全身图像的全自动分割

文献速递：肿瘤分割----基于卷积神经网络的系统，用于前列腺癌[68Ga]Ga-PSMA PET全身图像的全自动分割

01

文献速递介绍

前列腺特异性膜抗原（PSMA）PET/CT成像近年来在前列腺癌检测领域中获得了显著的重视。PSMA是一种在前列腺上皮和其他器官（如肾小管和唾液腺）中发现的膜蛋白。其表达在前列腺癌以及其他多种癌症类型中均有提高。PSMA表达水平与前列腺癌的阶段和等级相关。在PET/CT成像中，使用PSMA配体与细胞表面的PSMA受体结合，然后被细胞内化。有不同的PSMA配体可供选择，标记有镓-68（68Ga）或氟-18（18F），每种具有略微不同的特性。这些配体的例子包括[68Ga]Ga-PSMA-HBED-CC（[68Ga]Ga-PSMA-11）、[68Ga]Ga-PSMA-617、[68Ga]Ga-PSMAI&T、[18F]DCFBC、[18F]DCFPyL和[18F]F-PSMA-1007 。在这些中，[68Ga]Ga-PSMA-11已被广泛研究并用于不同阶段的前列腺癌成像。PSMA PET/CT成像被广泛接受用于检测复发前列腺癌的部位，与传统成像技术相比提供了更优越的检测能力。此外，越来越多的证据支持其在分期高危初级前列腺癌中的有效性。PSMA PET/CT还可以帮助确定放射配体治疗的资格，例如[177Lu]Lu-PSMA-617，其显示出高效率和低毒性，并可能监测晚期前列腺癌的治疗效果。检测和定位转移性前列腺癌（mPCa）病变对于靶向治疗程序（如放射疗法）至关重要。PET成像还可以帮助提取与PSMA表达相关的定量成像生物标志物，如肿瘤的最大标准化摄取值（SUVmax）、全身PSMA肿瘤体积（PSMA-TV）和全身总病变PSMA（TL-PSMA），这些可能具有预后潜力。此外，在影像组学领域，精确的疾病定位至关重要，放射组学通过分析大量成像特征来量化肿瘤特性，并指导对患者管理的个性化方法。影像组学在检测mPCa病变、预测未来转移的发展以及识别与晚期PCa患者总生存期（OS）相关的生物标志物方面表现出前景。然而，手动病变分割存在局限性，如劳动密集型，特别是对于肿瘤负担高的患者，且容易受到观察者变异性的影响。因此，完全自动化的疾病检测和分割将是非常有利的，因为它减少了手动用户输入，并加快了定量特征的提取，从而促进了更个性化的患者干预措施。

Title

题目

A convolutional neural network–based system for fully automatic segmentation of whole‑body [ 68Ga]Ga‑PSMA PET images in prostate cancer

基于卷积神经网络的系统，用于前列腺癌[68Ga]Ga-PSMA PET全身图像的全自动分割

Abstract

摘要

Purpose The aim of this study was development and evaluation of a fully automated tool for the detection and segmentation of mPCa lesions in whole-body [68Ga]Ga-PSMA-11 PET scans by using a nnU-Net framework.

目的：本研究的目的是开发和评估一个全自动工具，用于在全身[68Ga]Ga-PSMA-11 PET扫描中检测和分割转移性前列腺癌（mPCa）病变，该工具使用的是nnU-Net框架。

Results

结果

In terms of patient-level classifcation, the model achieved an accuracy of 83%, sensitivity of 92%, PPV of 77%, and NPV of 91% for the internal testing set. For lesion-level detection, the model achieved an accuracy of 87–94%, sensitivity of 88–95%, PPV of 98–100%, and F1-score of 93–97% for all testing sets. For voxel-level segmentation, the automated method achieved average values of 65–70% for DSC, 72–79% for PPV, 53–58% for IoU, and 62–73% for sensitivity in all testing sets. In the evaluation of volumetric parameters, there was a strong correlation between the manual and automated measurements of PSMA-TV and TL-PSMA for all centers.

在患者级分类方面，模型在内部测试集中达到了83%的准确度、92%的敏感性、77%的阳性预测值（PPV）和91%的阴性预测值（NPV）。在病变级检测方面，模型在所有测试集中的准确度为87-94%，敏感性为88-95%，阳性预测值为98-100%，F1分数为93-97%。在体素级分割方面，自动方法在所有测试集中的平均Dice相似系数（DSC）为65-70%，阳性预测值为72-79%，交并比（IoU）为53-58%，敏感性为62-73%。在体积参数的评估中，所有中心的PSMA-TV和TL-PSMA的手动和自动测量之间存在强相关性。

Methods

方法

In this multicenter study, a cohort of 412 patients from three diferent center with all indication of PCa who underwent [68Ga]Ga-PSMA-11 PET/CT were enrolled. Two hundred cases of center 1 dataset were used for training the model. A fully 3D convolutional neural network (CNN) is proposed which is based on the self-confguring nnU-Net framework. A subset of center 1 dataset and cases of center 2 and center 3 were used for testing of model. The performance of the segmentation pipeline that was developed was evaluated by comparing the fully automatic segmentation mask with the manual segmentation of the corresponding internal and external test sets in three levels including patient-level scan classifcation, lesion-level detection, and voxel-level segmentation. In addition, for comparison of PET-derived quantitative biomarkers between automated and manual segmentation, whole-body PSMA tumor volume (PSMA-TV) and total lesions PSMA uptake (TL-PSMA) were calculated.

在这项多中心研究中，共有412名来自三个不同中心的前列腺癌（PCa）患者参与，这些患者均接受了[68Ga]Ga-PSMA-11 PET/CT检查。中心1的200个病例数据用于训练模型。提出了一种完全基于自配置nnU-Net框架的3D卷积神经网络（CNN）。中心1的一部分数据集和中心2及中心3的病例用于模型测试。通过比较完全自动分割掩码与相应的内部和外部测试集的手动分割，从三个层面评估了所开发的分割流程的性能，包括患者级扫描分类、病变级检测和体素级分割。此外，为了比较自动分割和手动分割之间PET衍生的定量生物标志物，计算了全身PSMA肿瘤体积（PSMA-TV）和总病变PSMA摄取（TL-PSMA）。

Conclusions

结论

The deep learning networks presented here ofer promising solutions for automatically segmenting malignant lesions in prostate cancer patients using [68Ga]Ga-PSMA PET. These networks achieve a high level of accuracy in wholebody segmentation, as measured by the DSC and PPV at the voxel level. The resulting segmentations can be used for extraction of PET-derived quantitative biomarkers and utilized for treatment response assessment and radiomic studies. Keywords Prostate cancer · [68Ga]Ga-PSMA · PET/CT · Deep learning · Artifcial intelligence

这里展示的深度学习网络为使用[68Ga]Ga-PSMA PET自动分割前列腺癌患者中的恶性病变提供了有希望的解决方案。这些网络在整体分割中达到了高水平的准确度，以体素级的Dice相似系数（DSC）和阳性预测值（PPV）来衡量。得到的分割结果可用于提取PET衍生的定量生物标志物，并用于治疗反应评估和放射组学研究。关键词：前列腺癌、[68Ga]Ga-PSMA、PET/CT、深度学习、人工智能。

Figure

图

Fig. 1 An example of a whole body [68Ga]Ga-PSMA PET scan acquired from the skull to the middle of the thigh. The left image (A) shows a coronal view without any segmentation, the middle image (B) shows manual segmentations, and the right image © shows automated segmentations. The results showed good performance of CNN with Dice score of 85% and PPV of 85% in voxel-level segmentation

图1. 从头骨到大腿中部获取的全身[68Ga]Ga-PSMA PET扫描示例。左图（A）展示了未进行任何分割的冠状视图，中间图（B）展示了手动分割，右图（C）展示了自动分割。结果显示，在体素级分割中，CNN的性能良好，Dice分数为85%，阳性预测值（PPV）为85%。

Fig. 2 An example of a whole body [68Ga]Ga-PSMA PET scan acquired from the skull to the middle of the thigh. The left image (A) shows a coronal and axial views without any segmentation, the middle image (B) shows manual segmentations (green area), and the right image © shows automated segmentations. The results showed false negative in prostate bed and worse performance of CNN with Dice score of 0% and PPV of 0% in voxel-level segmentation

图2. 从头骨到大腿中部获取的全身[68Ga]Ga-PSMA PET扫描示例。左图（A）展示了未进行任何分割的冠状和轴向视图，中间图（B）展示了手动分割（绿色区域），右图（C）展示了自动分割。结果显示在前列腺床处出现假阴性，并且CNN在体素级分割中的性能较差，Dice分数为0%，阳性预测值（PPV）为0%。

Fig. 3 An example of a whole body [68Ga]Ga-PSMA PET scan acquired from the skull to the middle of the thigh with normal scan. The left image (A) shows a coronal and axial views without any segmentation, the middle image (B) shows manual segmentations, and the right image © shows automated segmentations. The results showed false positive in both breast nipples (red areas) and worse performance of CNN

图3. 从头骨到大腿中部获取的全身[68Ga]Ga-PSMA PET扫描正常扫描示例。左图（A）展示了未进行任何分割的冠状和轴向视图，中间图（B）展示了手动分割，右图（C）展示了自动分割。结果显示在两个乳头（红色区域）处出现假阳性，并且CNN的性能更差。

Fig. 4 Plots depicting the alteration in calculated metrics as the true-positive threshold is adjusted for the tasks of scan malignancy classifcation (A) and individual lesion detection for center 1 (internal testing set) (B), center 2 ©, and center 3 (D)

图 4 描述了在调整真阳性阈值时，用于扫描恶性分类任务（A）和中心 1（内部测试集）的单个病变检测（B）、中心 2（C）以及中心 3（D）的计算指标的变化情况。

Fig. 5 Scatter correlation plots for comparison of the manual and automated calculated PSMA-TV on the test sets. Strong positive correlations were observed between the automated and manually derived values

图 5 用于比较测试集上手动和自动计算的 PSMA-TV 的散点相关图。在自动和手动导出的值之间观察到强正相关。

Fig. 6 Scatter correlation plots for comparison of the manual and automated calculated TL-PSMA on the test sets. Strong positive correlations were observed between the automated and manually derived values

图 6 用于比较测试集上手动和自动计算的 TL-PSMA 的散点相关图。在自动和手动导出的值之间观察到强正相关。

Table

表

Table 1 The characteristics of dataset. There was no signifcant diference between group for all characteristics (p>0.05)

表1 数据集的特征。所有特征的组间差异均不显著（p>0.05）。

Table 2 Fully automated model performance calculated on the dedicated test sets of three centers

表2 在三个中心专用测试集上计算的全自动模型性能。

Table 3 The results of the manual and automatically calculated TL-PSMA and PSMA-TV. p<0.05 was considered as statistically signifcant

表 3 手动和自动计算的 TL-PSMA 和 PSMA-TV 结果。p<0.05 被认为是具有统计学意义的。