Abstract：
BACKGROUND: Accurate classification of gliomas is critical to the selection of immunotherapy, and MRI contains a large number of radiomic features that may suggest some prognostic relevant signals. We aim to predict new subtypes of gliomas using radiomic features and characterize their survival, immune, genomic profiles and drug response.
METHODS: We initially obtained 341 images of 36 patients from the CPTAC dataset for the development of deep learning models. Further 1812 images of 111 patients from TCGA_GBM and 152 images of 53 patients from TCGA_LGG were collected for testing and validation. A deep learning method based on Mask R-CNN was developed to identify new subtypes of glioma patients and compared the survival status, immune infiltration patterns, genomic signatures, specific drugs, and predictive models of different subtypes.
RESULTS: 200 glioma patients (mean age, 33 years ± 19 [standard deviation]) were enrolled. The accuracy of the deep learning model for identifying tumor regions achieved 88.3 % (98/111) in the test set and 83 % (44/53) in the validation set. The sample was divided into two subtypes based on radiomic features showed different prognostic outcomes (hazard ratio, 2.70). According to the results of the immune infiltration analysis, the subtype with a poorer prognosis was defined as the immunosilencing radiomic (ISR) subtype (n = 43), and the other subtype was the immunoactivated radiomic (IAR) subtype (n = 53). Subtype-specific genomic signatures distinguished celllines into ISR celllines (n = 9) and control celllines (n = 13), and identified eight ISR-specific drugs, four of which were validated by the OCTAD database. Three machine learning-based classifiers showed that radiomic and genomic co-features better predicted the radiomic subtypes of gliomas.
CONCLUSIONS: These findings provide insights into how radiogenomic could identify specific subtypes that predict prognosis, immune and drug sensitivity in a non-invasive manner.
METHODS: We initially obtained 341 images of 36 patients from the CPTAC dataset for the development of deep learning models. Further 1812 images of 111 patients from TCGA_GBM and 152 images of 53 patients from TCGA_LGG were collected for testing and validation. A deep learning method based on Mask R-CNN was developed to identify new subtypes of glioma patients and compared the survival status, immune infiltration patterns, genomic signatures, specific drugs, and predictive models of different subtypes.
RESULTS: 200 glioma patients (mean age, 33 years ± 19 [standard deviation]) were enrolled. The accuracy of the deep learning model for identifying tumor regions achieved 88.3 % (98/111) in the test set and 83 % (44/53) in the validation set. The sample was divided into two subtypes based on radiomic features showed different prognostic outcomes (hazard ratio, 2.70). According to the results of the immune infiltration analysis, the subtype with a poorer prognosis was defined as the immunosilencing radiomic (ISR) subtype (n = 43), and the other subtype was the immunoactivated radiomic (IAR) subtype (n = 53). Subtype-specific genomic signatures distinguished celllines into ISR celllines (n = 9) and control celllines (n = 13), and identified eight ISR-specific drugs, four of which were validated by the OCTAD database. Three machine learning-based classifiers showed that radiomic and genomic co-features better predicted the radiomic subtypes of gliomas.
CONCLUSIONS: These findings provide insights into how radiogenomic could identify specific subtypes that predict prognosis, immune and drug sensitivity in a non-invasive manner.
摘要：
背景：神经胶质瘤的准确分类对于选择免疫治疗至关重要，MRI包含大量的影像学特征，可能提示一些预后相关信号。我们的目标是预测新的亚型的胶质瘤使用影像学特征和表征他们的生存，免疫，基因组谱和药物反应。
方法：我们最初从CPTAC数据集中获得了36名患者的341张图像，用于开发深度学习模型。收集来自TCGA_GBM的111名患者的1812张图像和来自TCGA_LGG的53名患者的152张图像用于测试和验证。开发了一种基于MaskR-CNN的深度学习方法，用于识别胶质瘤患者的新亚型并比较其生存状态，免疫浸润模式，基因组特征，特定药物,和不同亚型的预测模型。
结果：200例胶质瘤患者（平均年龄，33年±19[标准偏差])。用于识别肿瘤区域的深度学习模型的准确度在测试集中达到88.3%(98/111)，在验证集中达到83%(44/53)。根据放射学特征将样本分为两个亚型，显示出不同的预后结果(风险比，2.70).根据免疫浸润分析结果,预后较差的亚型被定义为免疫沉默放射组学（ISR）亚型（n=43），另一种亚型是免疫激活的放射组学（IAR）亚型（n=53）。亚型特异性基因组特征将细胞系分为ISR细胞系（n=9）和对照细胞系（n=13）。并鉴定了八种ISR特异性药物，其中4个已通过OCTAD数据库验证.三个基于机器学习的分类器显示放射学和基因组共同特征更好地预测胶质瘤的放射学亚型。
结论：这些发现提供了关于放射基因组如何识别预测预后的特定亚型的见解。非侵入性的免疫和药物敏感性。
方法：我们最初从CPTAC数据集中获得了36名患者的341张图像，用于开发深度学习模型。收集来自TCGA_GBM的111名患者的1812张图像和来自TCGA_LGG的53名患者的152张图像用于测试和验证。开发了一种基于MaskR-CNN的深度学习方法，用于识别胶质瘤患者的新亚型并比较其生存状态，免疫浸润模式，基因组特征，特定药物,和不同亚型的预测模型。
结果：200例胶质瘤患者（平均年龄，33年±19[标准偏差])。用于识别肿瘤区域的深度学习模型的准确度在测试集中达到88.3%(98/111)，在验证集中达到83%(44/53)。根据放射学特征将样本分为两个亚型，显示出不同的预后结果(风险比，2.70).根据免疫浸润分析结果,预后较差的亚型被定义为免疫沉默放射组学（ISR）亚型（n=43），另一种亚型是免疫激活的放射组学（IAR）亚型（n=53）。亚型特异性基因组特征将细胞系分为ISR细胞系（n=9）和对照细胞系（n=13）。并鉴定了八种ISR特异性药物，其中4个已通过OCTAD数据库验证.三个基于机器学习的分类器显示放射学和基因组共同特征更好地预测胶质瘤的放射学亚型。
结论：这些发现提供了关于放射基因组如何识别预测预后的特定亚型的见解。非侵入性的免疫和药物敏感性。
