| 摘要: | ?管近年?肺癌治?取得了一些?展，但患者的五年生存率??15%。因此，??影像?已被用作??非小?胞肺癌（NSCLC）患者生存率的工具。本研究的目?是通??算机???描（CT?描）??放射?特征?名（??放射?和深度放射?），以??NSCLC患者的生存率。我??The Cancer Imaging Archive（TCIA）中已?布的??NSCLC?据集（NSCLC-Radiomics和NSCLC-Radiogenomics）?行了回?性分析。通?特征??和降?的??步?，??了放射?特征?名。
在初始?段，我??估了??放射?的有效性。研究?果?示，利用??放射?特征?名的模型在??NSCLC患者的?体生存率方面表?出相?的?力。
考?到?CT?描中存在一???放射???用于不同癌症?瘤生存??的可能性，我?一直在不?改?和利用?一集合的有效性，以??不同?型的?性?瘤生存率。本研究?涵?了??、??癌等癌症部位。我??2019年??和???瘤分割??（KiTS19）和The Cancer Imaging Archive中的??部???胞癌（HNSCC）?据集添加到我?的研究中。研究?果表明，???放射?特征?名和?床因素?合的?合模型，在各种?性?瘤的生存??背景下，比?依?于??放射?特征?名的模型具有更好的??能力。在Lung 1??集和Lung 2??集中，?合模型的iAUC分??0.621（95% CI: 0.588,0.654）和0.736（95% CI: 0.645,0.819）。??部和????集的?合模型的iAUC分??0.732（95% CI: 0.655,0.809）和0.834（95% CI: 0.722,0.946）。
最后，本研究?旨在?建一种深度??方法，利用?床、深度放射?特征和??放射?特征???NSCLC患者的?体生存。我?采用三?（3D）卷?神?网?（CNN）存活深度神?网?架构提取深度放射?特征，并??非小?胞肺癌（NSCLC）患者的?体生存。?深度放射?特征与??放射?特征和?床??合并。模型的有效性使用一致性指?（C-index）?行?估。我?的研究得出??，通?深度????床、深度放射?和??放射?特征整合起?，能?准确??NSCLC患者的?体生存。?合模型（使用?床、深度放射?和??放射?等3???）?用Deepsurv方法在与其他模型比????了最高的效率（Lung 1??集C-index?0.733，Lung 2??集C-index?0.751)
In spite of advancements made in lung cancer treatment in recent times, the survival rate for patients at the 5-year mark remains a mere 15%. Therefore, medical imaging has been used as a tool for predicting survival rates in patients with non-small cell lung cancer (NSCLC). The objective of this study was to develop radiomics feature signatures (traditional-radiomics and deep-radiomics) from computed tomography (CT) scans to predict survival rates in NSCLC patients. We conducted a retrospective analysis of two datasets of NSCLC (NSCLC-Radiomics and NSCLC-Radiogenomics) that were published in The Cancer Imaging Archive (TCIA). The radiomics signatures are found by the statistical steps of features selection, reducing the features dimension.
In the initial phase, the effectiveness of traditional-radiomics was assessed. The findings revealed that models utilizing traditional-radiomics signatures demonstrated considerable promise in predicting the overall survival of individuals with NSCLC.
By considering the possibility of existing a set of traditional-radiomics markers from CT scans for diverse cancer tumors survival prediction, there has been ongoing progress in enhancing and utilizing the effectiveness of this set in forecasting survival rates for different types of malignancies. This research encompassed cancer sites such as the kidney, as well as head and neck cancer. Two data sets (the 2019 Kidney and Kidney Tumor Segmentation Competition (KiTS19) and Head and neck cancer squamous cell carcinoma (HNSCC) in the TCIA archive) were added to our study. The findings indicated that a combined model incorporating both traditional-radiomics and clinical factors demonstrated superior predictive capability compared to relying solely on traditional-radiomics in the context of survival prediction in various malignancies. The Integrated area under curve (iAUC) of the combined model in Lung 1 training set, Lung 2 testing set is 0.621 (95% CI: 0.588,0.654) and 0.736 (95% CI: 0.645,0.819), respectively. Head & neck (H&N) and Kidney validation set obtained combined model’s iAUC is 0.732 (95% CI: 0.655,0.809) and 0.834 (95% CI: 0.722,0.946), respectively.
Lastly, the study also aimed to create a deep learning approach that uses clinical, deep-radiomics features, and traditional-radiomics features to predict overall survival in NSCLC patients. We utilized a 3-dimensional (3D) convolutional neural network (CNN) survival deep neural network architecture to extract deep-radiomics features and predict the overall survival of patients with NSCLC. The deep-radiomics features were merged with traditional-radiomics signatures and clinical parameters. The model's effectiveness was evaluated using the concordance index (C-index). Our study concluded that integrating clinical, deep-radiomics, and traditional-radiomics features through deep learning enabled accurate prediction of overall survival in NSCLC patients. Combined model (using 3 parameters including clinical, deep-radiomics and traditional-radiomics) applied Deepsurv method achieve the highest efficiency when compared with other models (C index is 0.733 in Lung 1 training set, C index is 0.751 in Lung 2 testing set). |
