Geles: A Novel Imaging Informatics System for Generalizable Lesion Identification in Neuroendocrine Tumors

项目摘要 胃肠胰腺神经内分泌肿瘤（GEP-NETs）是一种很难检测的肿瘤， 目前在先进的阶段，与肝脏作为最常见的部位转移。68 Ga和64 Cu DOTATATE 正电子发射断层扫描-计算机断层扫描（PET/CT）是识别 生长抑素受体亚型2阳性GEP-NETs和靶向肽放射性核素受体治疗， 177 Lu DOTATATE是许多患者最有效的全身治疗方法。尽管有明显的优势， 与既往标准治疗相比，绝大多数患者（99%）的无进展生存期 完全缓解，需要额外治疗。治疗的进一步发展需要准确的 评估对治疗的反应。然而，目前存在对自动化的未满足的医疗需求， 68 Ga DOTATATE阳性疾病负担的标准化量化，这可能对 新的治疗药物方案开发。基于深度学习的方法最近已被应用于 自动病变检测和量化，并已达到最先进的性能。这些方法， 但是，不要考虑训练数据和测试数据之间的数据集/域偏移。在数据集/域转移中，数据 用于构建和训练模型的分布可能与用于模型测试的分布显著不同。 因此，不考虑域转移的模型将不能很好地推广到看不见的数据，导致较差的 病变检测性能。在这项拟议的研究中，我们将开发一种新的基于深度学习的成像技术， 信息学系统，称为Geles，用于GEP-NET PET/CT中肝脏的自动化、可推广病变检测 显像该系统将使用列表模式数据采集来产生一个大型的、多样化的带注释的训练数据集， 其次是新的对抗性学习，以增强模型的泛化能力。 建议的Geles系统将包括两个模块，领域泛化和领域自适应。要求1 将开发一个对抗域泛化模块，该模块可泛化到看不见的域或资源。 本模块将构建一个具有领域对抗学习的深度神经网络，并提取领域不变 用于单个病变识别的特征表示，以便系统可以推广到看不见的领域 数据，例如来自不同机构、设备、成像协议和其他变体的PET图像。目标2将 开发一个面向目标的领域适应模块，自动适应新的特定数据集， 兴趣（即，目标数据集）。给定来自某个目标数据集的一小组未注释图像，该模块 将进行低资源无监督域自适应，以进一步提高病变检测性能。 具体来说，它将建立一个新的，增强的生成对抗网络，用于图像到图像的翻译， 低资源设置，使Geles可以利用有限的，未注释的具体目标数据和行为 以目标为导向，增强的病变检测。

PROJECT SUMMARY Gastroenteropancreatic neuroendocrine tumors (GEP-NETs) are difficult to detect tumors which commonly present at advanced stages, with the liver as the most common site of metastases. 68Ga and 64Cu DOTATATE positron emission tomography-computed tomography (PET/CT) are the most sensitive methods to identify somatostatin receptor subtype 2 positive GEP-NETs, and targeted peptide radionuclide receptor therapy with 177Lu DOTATATE is the most effective systemic therapy for many patients. Despite the clear advantage in progression-free survival compared to prior standard of care, the vast majority of patients (99%) do not have complete response and require additional therapies. Further development of treatments requires an accurate assessment of the response to therapy. However, there is currently an unmet medical need for automated, standardized quantification of 68Ga DOTATATE positive disease burden, which could have a great impact on novel therapeutic drug regimen development. Deep learning-based approaches have recently been applied to automated lesion detection and quantification, and have achieved state-of-the-art performance. These methods, however, do not consider dataset/domain shifts between training and testing data. In dataset/domain shifts, data used to build and train models might have a significantly different distribution from that used for model testing. Therefore, models without considering domain shifts would not generalize well to unseen data, leading to poor lesion detection performance. In this proposed research, we will develop a novel deep learning-based imaging informatics system, termed Geles, for automated, Generalizable lesion detection for livers in GEP-NET PET/CT imaging. This system will use list-mode data acquisition to produce a large, diverse annotated training dataset, followed by novel adversarial learning to enhance model generalizability. The proposed Geles system will consist of two modules, domain generalization and domain adaptation. Aim 1 will develop an adversarial domain generalization module that is generalizable to unseen domains or resources. This module will build a deep neural network with domain-adversarial learning and extract domain-invariant feature representations for individual lesion identification, so that the system can generalize to unseen domain data, such as PET images from different institutions, devices, imaging protocols, and other variations. Aim 2 will develop a target-oriented domain adaptation module that is automatically adaptable to new specific datasets of interest (i.e., target datasets). Given a small set of unannotated images from a certain target dataset, this module will conduct low-resource unsupervised domain adaptation to further boost the lesion detection performance. Specifically, it will build a novel, augmented generative adversarial network for image-to-image translation in a low-resource setting, so that Geles can take advantage of limited, unannotated specific target data and conduct target-oriented, enhanced lesion detection.