免疫检查点抑制剂（ICIs）在NSCLC治疗中取得突破性进展，但ICIs治疗个体差异大、总体响应率偏低是目前应用的核心问题，目前通过活检获得的PD-L1等生物筛选指标存在局限性。深度学习是一种数据驱动的研究方法，可从相对易获取的影像、液体活检以及临床数据中挖掘信息，图神经网络较目前用于深度学习的标准神经网络在推理非结构化数据等方面有较大优势。我们前期研究显示FDG PET/CT代谢参数与ICIs新辅助治疗后NSCLC的病理缓解率显著相关，提示其在反映NSCLC免疫检查点相关生物学特性方面具有较大优势。据此我们拟分析不同ICIs新辅助治疗预后的NSCLC患者PET/CT组学特征和代谢参数差异，筛选有预测价值的参数并与其免疫治疗相关生物指标关联，探讨其分子基础；同时将上述参数与液体活检、临床数据联合，建立基于PET/CT图神经网络的NSCLC免疫新辅助治疗获益人群的模型并行临床验证。

Immune checkpoint inhibitors (ICIs) have progressed tremendously by the clinical outcomes of non-small-cell lung cancer (NSCLC) patients. However, the great individual difference on objective response is the key issue of clinical application till now, and there are limitations in biological screening markers such as PD-L1 obtained by biopsy. Deep learning is a data-driven research method, which can extract data from easy-get images, liquid biopsy, and clinical data. Graph neural network (GNN) has greater advantages on inferring unstructured data than that of the standard neural network for deep learning. Our previous study found that metabolic parameters of FDG PET/CT are significantly correlated with pathologic response rate of NSCLC after ICIs neoadjuvant therapy, and suggested that these parameters may have some advantages on reflecting immune microenvironment of NSCLC. Accordingly, we further propose to find meaningful parameters (including radiomics indexes and metabolic parameters) of PET-CT in different prognosis of resectable NSCLC patients who will accept ICIs neoadjuvant therapy, explore the correlation with the expression of biological immunotherapy biomarkers, and discuss the FDG PET/CT role in reflecting prognosis of NSCLC under immunotherapy on molecular level. At the same time, we will establish and validate a predictive model, which can forecast the benefits to patients from NSCLC immunotherapy, by using GNN algorithm to deeply learn the unstructured data (including the above meaningful PET-CT parameters, liquid biopsy and clinical data).