Li Bo, Su Jie, Liu Kai, Hu Chunfeng
Department of Radiology, Affiliated Hospital of Xuzhou Medical University, Xuzhou Medical University, Xuzhou, China.
Eur J Radiol Open. 2024 Jan 19;12:100549. doi: 10.1016/j.ejro.2024.100549. eCollection 2024 Jun.
Programmed cell death protein-1 ligand (PD-L1) is an important prognostic predictor for immunotherapy of non-small cell lung cancer (NSCLC). This study aimed to develop a non-invasive deep learning and radiomics model based on positron emission tomography and computed tomography (PET/CT) to predict PD-L1 expression in NSCLC.
A total of 136 patients with NSCLC between January 2021 and September 2022 were enrolled in this study. The patients were randomly divided into the training dataset and the validation dataset in a ratio of 7:3. Radiomics feature and deep learning feature were extracted from their PET/CT images. The Mann-whitney U-test, Least Absolute Shrinkage and Selection Operator algorithm and Spearman correlation analysis were used to select the top significant features. Then we developed a radiomics model, a deep learning model, and a fusion model based on the selected features. The performance of three models were compared by the area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value, and negative predictive value.
Of the patients, 42 patients were PD-L1 negative and 94 patients were PD-L1 positive. A total of 2446 radiomics features and 4096 deep learning features were extracted per patient. In the training dataset, the fusion model achieved a highest AUC (0.954, 95% confident internal [CI]: 0.890-0.986) compared with the radiomics model (0.829, 95%CI: 0.738-0.898) and the deep learning model (0.935, 95%CI: 0.865-0.975). In the validation dataset, the AUC of the fusion model (0.910, 95% CI: 0.779-0.977) was also higher than that of the radiomics model (0.785, 95% CI: 0.628-0.897) and the deep learning model (0.867, 95% CI: 0.724-0.952).
The PET/CT-based deep learning radiomics model can predict the PD-L1 expression accurately in NSCLC patients, and provides a non-invasive tool for clinicians to select positive PD-L1 patients.
程序性细胞死亡蛋白1配体（PD-L1）是非小细胞肺癌（NSCLC）免疫治疗的重要预后预测指标。本研究旨在基于正电子发射断层扫描和计算机断层扫描（PET/CT）开发一种非侵入性深度学习和放射组学模型，以预测NSCLC中的PD-L1表达。
本研究纳入了2021年1月至2022年9月期间的136例NSCLC患者。患者按7:3的比例随机分为训练数据集和验证数据集。从他们的PET/CT图像中提取放射组学特征和深度学习特征。使用曼-惠特尼U检验、最小绝对收缩和选择算子算法以及斯皮尔曼相关性分析来选择最显著的特征。然后，我们基于所选特征开发了一个放射组学模型、一个深度学习模型和一个融合模型。通过曲线下面积（AUC）、敏感性、特异性、准确性、阳性预测值和阴性预测值比较了三种模型的性能。
在这些患者中，42例患者PD-L1阴性，94例患者PD-L1阳性。每位患者共提取了2446个放射组学特征和4096个深度学习特征。在训练数据集中，与放射组学模型（AUC = 0.829，95%可信区间[CI]：0.738 - 0.898）和深度学习模型（AUC = 0.935，95%CI：0.865 - 0.975）相比，融合模型获得了最高的AUC（0.954，95%CI：0.890 - 0.986）。在验证数据集中，融合模型的AUC（0.910，95%CI：0.779 - 0.977）也高于放射组学模型（0.785，95%CI：0.628 - 0.897）和深度学习模型（0.867，95%CI：0.724 - 0.952）。
基于PET/CT的深度学习放射组学模型可以准确预测NSCLC患者的PD-L1表达，并为临床医生选择PD-L1阳性患者提供了一种非侵入性工具。
