Beaumont Hubert, Iannessi Antoine, Bertrand Anne-Sophie, Cucchi Jean Michel, Lucidarme Olivier
Median Technologies, 06560, Valbonne, France.
Centre Antoine Lacassagne, 06100, Nice, France.
Eur Radiol. 2021 Aug;31(8):6059-6068. doi: 10.1007/s00330-020-07641-8. Epub 2021 Jan 18.
Following the craze for radiomic features (RF), their lack of reliability raised the question of the generalizability of classification models. Inter-site harmonization of images therefore becomes a central issue. We compared RF harmonization processing designed to detect liver diseases in CT images.
We retrospectively analyzed 76 multi-center portal CT series of non-diseased (NDL) and diseased liver (DL) patients. In each series, we positioned volumes of interest in spleen and liver, then extracted 9 RF (histogram and texture). We evaluated two RF harmonization approaches. First, in each series, we computed the Z-score of liver measurements based on those computed in the spleen. Second, we evaluated the ComBat method according to each imaging center; parameters were computed in the spleen and applied to the liver. We compared RF distributions and classification performances before/after harmonization. We classified NDL versus spleen and versus DL tissues.
The RF distributions were all different between liver and spleen (p < 0.05). The Z-score harmonization outperformed for the detection of liver versus spleen: AUC = 93.1% (p < 0.001). For the detection of DL versus NDL, in a case/control setting, we found no differences between the harmonizations: mean AUC = 73.6% (p = 0.49). Using the whole datasets, the performances were improved using ComBat (p = 0.05) AUC = 82.4% and degraded with Z-score AUC = 67.4% (p = 0.008).
Data harmonization requires to first focus on data structuring to not degrade the performances of subsequent classifications. Liver tissue classification after harmonization of spleen-based RF is a promising strategy for improving the detection of DL tissue.
• Variability of acquisition parameter makes radiomics of CT features non-reproducible. • Data harmonization can help circumvent the inter-site variability of acquisition protocols. • Inter-site harmonization must be carefully implemented and requires designing consistent data sets.
随着对放射组学特征（RF）的狂热，其可靠性的缺乏引发了分类模型可推广性的问题。因此，图像的跨站点协调成为一个核心问题。我们比较了旨在检测CT图像中肝脏疾病的RF协调处理方法。
我们回顾性分析了76例非疾病（NDL）和患病肝脏（DL）患者的多中心门静脉CT系列。在每个系列中，我们在脾脏和肝脏中定位感兴趣的体积，然后提取9个RF（直方图和纹理）。我们评估了两种RF协调方法。首先，在每个系列中，我们根据在脾脏中计算的测量值计算肝脏测量值的Z分数。其次，我们根据每个成像中心评估ComBat方法；在脾脏中计算参数并应用于肝脏。我们比较了协调前后的RF分布和分类性能。我们将NDL与脾脏以及与DL组织进行分类。
肝脏和脾脏之间的RF分布均不同（p < 0.05）。Z分数协调在检测肝脏与脾脏方面表现更优：AUC = 93.1%（p < 0.001）。对于在病例/对照设置中检测DL与NDL，我们发现协调方法之间没有差异：平均AUC = 73.6%（p = 0.49）。使用整个数据集，使用ComBat时性能得到改善（p = 0.05），AUC = 82.4%，而使用Z分数时性能下降，AUC = 67.4%（p = 0.008）。
数据协调首先需要关注数据结构，以免降低后续分类的性能。基于脾脏的RF协调后的肝脏组织分类是改善DL组织检测的一种有前景的策略。
• 采集参数的变异性使得CT特征的放射组学不可重复。• 数据协调有助于规避采集协议的跨站点变异性。• 跨站点协调必须谨慎实施，并且需要设计一致的数据集。
