Semi-automatic classification of prostate cancer on multi-parametric MR imaging using a multi-channel 3D convolutional neural network
Semi-automatic classification of prostate cancer on multi-parametric MR imaging using a multi-channel 3D convolutional neural network
复制标题
DOI：
10.1007/s00330-019-06417-z
复制
发表时间：
2020-02-01
影响因子：
5.9
通讯作者：
Penzkofer, Tobias
中科院分区：
文献类型：
--
作者：
Aldoj, Nader;Lukas, Steffen;Penzkofer, Tobias
目的 提出一种基于深度学习的方法，使用 3D 卷积神经网络 (CNN) 基于多参数磁共振 (MR) 成像进行半自动前列腺癌分类。方法对 200 例患者共 318 个具有组织学相关性的病变进行分析。使用不同 MRI 序列的不同组合作为输入（例如 T2 加权、表观扩散系数 (ADC)、扩散加权图像和 K-trans）设计、训练和验证一种新颖的 CNN，并测试和讨论了不同序列对网络性能的影响。通过测试所有相关数据组合来证明建模方法的特定选择是合理的。使用八重交叉验证对该模型进行训练和验证。结果 在以活检结果为参考标准的显着前列腺癌检测方面，3D CNN 实现的受试者工作特征曲线下面积 (AUC) 范围为 0.89（敏感性和特异性分别为 88.6% 和 90.0%）至 0.91（敏感性和特异性分别为 81.2% 和 90.5%），ADC、DWI 和 K-trans 的平均 AUC 为 0.897输入组合。其他组合在整体性能和平均 AUC 方面得分较低，其中使用 T2w 和 K-trans 时性能差异显着，p 值为 0.02；使用 T2w、ADC 和 DWI 时为 0.00025。因此，前列腺癌分类性能与经验丰富的放射科医生使用前列腺成像报告和数据系统 (PI-RADS) 报告的性能相当。病变大小和最大直径对网络性能没有影响。结论 3D CNN 在检测有临床意义的前列腺癌方面的诊断性能具有良好的 AUC 和敏感性和高特异性。
Objective To present a deep learning-based approach for semi-automatic prostate cancer classification based on multi-parametric magnetic resonance (MR) imaging using a 3D convolutional neural network (CNN). Methods Two hundred patients with a total of 318 lesions for which histological correlation was available were analyzed. A novel CNN was designed, trained, and validated using different combinations of distinct MRI sequences as input (e.g., T2-weighted, apparent diffusion coefficient (ADC), diffusion-weighted images, and K-trans) and the effect of different sequences on the network's performance was tested and discussed. The particular choice of modeling approach was justified by testing all relevant data combinations. The model was trained and validated using eightfold cross-validation. Results In terms of detection of significant prostate cancer defined by biopsy results as the reference standard, the 3D CNN achieved an area under the curve (AUC) of the receiver operating characteristics ranging from 0.89 (88.6% and 90.0% for sensitivity and specificity respectively) to 0.91 (81.2% and 90.5% for sensitivity and specificity respectively) with an average AUC of 0.897 for the ADC, DWI, and K-trans input combination. The other combinations scored less in terms of overall performance and average AUC, where the difference in performance was significant with a p value of 0.02 when using T2w and K-trans; and 0.00025 when using T2w, ADC, and DWI. Prostate cancer classification performance is thus comparable to that reported for experienced radiologists using the prostate imaging reporting and data system (PI-RADS). Lesion size and largest diameter had no effect on the network's performance. Conclusion The diagnostic performance of the 3D CNN in detecting clinically significant prostate cancer is characterized by a good AUC and sensitivity and high specificity.