Radiology and Imaging Sciences, Clinical Center (CC), National Institutes of Health, Bethesda, MD, USA.
Artificial Intelligence Resource, Clinical Center, National Institutes of Health, Bethesda, MD, USA.
Abdom Radiol (NY). 2022 Oct;47(10):3554-3562. doi: 10.1007/s00261-022-03610-5. Epub 2022 Jul 22.
Upfront knowledge of tumor growth rates of clear cell renal cell carcinoma in von Hippel-Lindau syndrome (VHL) patients can allow for a more personalized approach to either surveillance imaging frequency or surgical planning. In this study, we implement a machine learning algorithm utilizing radiomic features of renal tumors identified on baseline magnetic resonance imaging (MRI) in VHL patients to predict the volumetric growth rate category of these tumors.
A total of 73 VHL patients with 173 pathologically confirmed Clear Cell Renal Cell Carcinoma (ccRCCs) underwent MRI at least at two different time points between 2015 and 2021. Each tumor was manually segmented in excretory phase contrast T1 weighed MRI and co-registered on pre-contrast, corticomedullary and nephrographic phases. Radiomic features and volumetric data from each tumor were extracted using the PyRadiomics library in Python (4544 total features). Tumor doubling time (DT) was calculated and patients were divided into two groups: DT < = 1 year and DT > 1 year. Random forest classifier (RFC) was used to predict the DT category. To measure prediction performance, the cohort was randomly divided into 100 training and test sets (80% and 20%). Model performance was evaluated using area under curve of receiver operating characteristic curve (AUC-ROC), as well as accuracy, F1, precision and recall, reported as percentages with 95% confidence intervals (CIs).
The average age of patients was 47.2 ± 10.3 years. Mean interval between MRIs for each patient was 1.3 years. Tumors included in this study were categorized into 155 Grade 2; 16 Grade 3; and 2 Grade 4. Mean accuracy of RFC model was 79.0% [67.4-90.6] and mean AUC-ROC of 0.795 [0.608-0.988]. The accuracy for predicting DT classes was not different among the MRI sequences (P-value = 0.56).
Here we demonstrate the utility of machine learning in accurately predicting the renal tumor growth rate category of VHL patients based on radiomic features extracted from different T1-weighted pre- and post-contrast MRI sequences.
了解 von Hippel-Lindau 综合征（VHL）患者透明细胞肾细胞癌（ccRCC）的肿瘤生长速度，有助于制定更个体化的随访影像学检查频率或手术计划。本研究利用 VHL 患者基线磁共振成像（MRI）上识别的肾肿瘤的放射组学特征，建立机器学习算法，预测这些肿瘤的容积增长率类别。
本研究共纳入 73 例 VHL 患者，这些患者共 173 个经病理证实的透明细胞肾细胞癌（ccRCC）。所有患者在 2015 年至 2021 年间均至少接受过两次不同时间点的 MRI 检查。在排泄期对比 T1 加权 MRI 上手动对每个肿瘤进行分割，并与预对比、皮质髓质和肾图期进行配准。使用 Python 中的 PyRadiomics 库（共提取 4544 个特征）提取肿瘤的放射组学特征和容积数据。计算肿瘤倍增时间（DT），并将患者分为两组：DT≤1 年和 DT＞1 年。使用随机森林分类器（RFC）预测 DT 类别。为了评估预测性能，将队列随机分为 100 个训练集和测试集（80%和 20%）。使用受试者工作特征曲线下面积（AUC-ROC）以及准确性、F1、精度和召回率来评估模型性能，以百分比表示，置信区间（CI）为 95%。
患者的平均年龄为 47.2±10.3 岁。每位患者两次 MRI 检查的平均间隔为 1.3 年。本研究纳入的肿瘤分为 155 个 2 级；16 个 3 级；和 2 个 4 级。RFC 模型的平均准确性为 79.0%[67.4-90.6]，平均 AUC-ROC 为 0.795[0.608-0.988]。不同 MRI 序列预测 DT 类别之间的准确性无差异（P 值=0.56）。
本研究基于不同 T1 加权对比前后 MRI 序列提取的放射组学特征，证明了机器学习在准确预测 VHL 患者肾肿瘤生长速度类别方面的应用价值。
