毕业生知识库

鼻咽癌是一种源于鼻咽粘膜的上皮组织恶性肿瘤，在我国华南地区高发，该病多见于青壮年，给我国社会和经济造成了巨大负担。鼻咽癌诊疗可以分为筛查、诊断、治疗三个阶段，但多数鼻咽癌患者诊断时已经发展为晚期，导致预后较差。因此，鼻咽癌诊断前的早期筛查和诊断后的预后预测对改善鼻咽癌患者的预后至关重要。

深度学习等人工智能技术可以将医学影像转化为高维定量特征，为临床筛查、诊断、预后预测和治疗决策提供支持，在鼻咽癌等多个癌种中已被证明具有潜在的临床应用价值。通常，深度学习算法需要大量数据的集中训练，但医学影像数据本身分散在各个数据中心，受限于数据安全、患者隐私保护，难以被集中在一个中心。随着联邦学习技术的发展，在多个离散独立的数据中心之间，不涉及中心之间数据交换得到前提下，已经可以完成深度学习模型的训练，并达到几乎和集中式训练相近的性能。

但是目前联邦学习尚未应用于鼻咽癌诊疗领域，并且鼻咽癌数据存在患者地域分布差异大的情况，给常规的联邦学习的应用带来了挑战。针对鼻咽癌的早期筛查和预后预测问题，本文开展了基于本地学习和联邦学习的鼻咽癌辅助诊疗算法研究，利用本地学习中各中心分布的差异特点来优化联邦学习模型，实现在中心间存在严重非独立同分布情况下的模型训练和验证，主要工作包含以下三部分：

（1）针对复发鼻咽癌患者的预后预测难的临床问题和现有临床预后预测模型效果有待提高的技术问题，本文基于复发鼻咽癌 PET/CT 影像提出了深度学习多模态融合预后预测模型，显著提高了复发鼻咽癌的预后预测效果。具体来说，本文以 EfficientNet 模型和 DeepSurv 损失函数为基础，构建了融合 PET 影像和CT 影像的深度学习多模态融合预后预测模型，以总生存期为目标，端到端地学习一组与预后相关的深度学习特征，再建模 Cox 生存期回归模型。实验表明，根据多模态融合模型的预测值，可以将患者划分为低危组和高危组，低危组比高危组的 3 年生存率获益（低危组与高危组间 3 年生存率差值）为 28.5%，相比常规临床模型（19.2%）提高了 9.3%，有望使这些患者选择更适合的治疗方式，改善患者预后。该研究表明，深度学习能够从肿瘤的影像数据中挖掘临床特征所不具有的肿瘤异质性信息，有望辅助复发鼻咽癌的治疗决策。

（2）针对鼻咽癌早期筛查难的临床问题和多中心鼻咽癌早筛数据隐私保护的技术问题，本文构建了用于鼻咽癌早期筛查的联邦学习训练框架，实现了符合数据安全规范的多中心鼻咽癌早筛模型的训练和验证。具体来说，本文以联邦平均策略为基础，在本地以多进程模拟的方式部署了联邦学习训练框架，验证了联邦学习模型的可用性。并且，针对多中心数据存在的非独立同分布问题，本文利用私有层联邦策略和联邦-微调策略优化联邦学习训练过程，从模型角度和训练角度两方面提升联邦学习在各个中心的效果，相较于本地训练的模型，实现了更准确的鼻咽癌筛查。实验表明，受数据非独立同分布影响，常规联邦平均策略仅能取得 0.8859 的分类 AUC，低于本地学习的 0.8926，经过优化后，联邦学习的分类 AUC 为 0.8962，接近于集中式学习的 0.9023。该研究表明，联邦学习能够在数据安全、隐私保护的限制下，利用多中心的鼻咽癌早筛数据，取得比单中心本地训练模型更优的早期筛查效果。

（3）针对多中心鼻咽癌早筛数据集的严重非独立同分布性，以及联邦学习早筛模型在多中心数据集上泛化性能差的问题，本文提出了基于模型更新距离的自适应数据重采样算法，实现了在严重非独立同分布数据上的联邦学习模型训练和验证。具体来说，该研究以联邦平均策略为基础，通过统计联邦学习训练过程中各中心上传的模型更新距离，以全局更新距离平均值为目标，让各个中心的更新距离向全局平均靠近，减少各中心更新距离的差异。在此基础上，定义了重采样系数的计算公式，让各中心按比例对数据集进行重采样，使得模型能够更多地学习小数据量、难分类数据中心的更新方向。实验表明，加入自适应重采样算法后，联邦学习的分类 AUC 提升至 0.9030，高于集中式学习。该研究表明，自适应数据重采样算法可以有效地缓解联邦学习中由于数据非独立同分布性导致的数据分布不均衡问题，提高联邦学习的分类效果和训练效率。

综上所述，本文围绕鼻咽癌预后预测和早期筛查的问题，在本地学习和联邦学习方面分别提出了深度学习多模态融合预后预测模型和基于模型更新距离自适应重采样算法，实现了基于本地学习和联邦学习的鼻咽癌辅助诊疗，为提高患者生存期提供了有效的工具。相关工作本人以第一作者（含共同第一作者）发表于临床主流期刊 European Journal of Nuclear Medicine and Molecular Imaging 和Therapeutic Advances in Medical Oncology，申请了一项国家发明专利并登记了两项软件著作权。

Nasopharyngeal carcinoma (NPC) is a malignant tumor originating from the epithelial tissue of the nasopharyngeal mucosa. It is highly prevalent in the southern regions of China and mainly affects young adults, causing a significant burden on the country’s society and economy. NPC diagnosis and treatment can be divided into three stages: screening, diagnosis, and treatment. However, most NPC patients are diagnosed in the advanced stages, resulting in poor prognosis. Therefore, early screening before NPC diagnosis and prognosis prediction after diagnosis are crucial for improving the prognosis of NPC patients.

Artificial intelligence technologies such as deep learning can transform medical images into high-dimensional quantitative features, providing support for clinical screening, diagnosis, prognosis prediction, and treatment decisions. These technologies have demonstrated potential clinical applications in various cancers, including NPC. Typically, deep learning algorithms require centralized training with a large amount of data, but medical image data is scattered across different data centers and is limited by data security and patient privacy protection, making it difficult to centralize. With the development of federated learning technology, deep learning models can be trained among multiple discrete independent data centers without involving data exchange between centers, achieving performance close to centralized training. However, federated learning has not yet been applied to NPC diagnosis and treatment, and the impact of differences in the geographic distribution of NPC data presents challenges for conventional federated learning applications.

To address the issue of early screening and prognosis prediction for NPC, this study conducted research on NPC assisted diagnosis and treatment algorithms based on local learning and federated learning. By leveraging the differences in the distribution of each center in local learning to optimize federated learning models, we achieved model training and application in situations with severe non-independent and identically distributed data in centers. The main work of this study includes the following three parts:

(1) In response to the clinical problem of difficult prognosis prediction for patients with recurrent nasopharyngeal carcinoma and the technical problem of poor performance of existing clinical prognosis prediction models, this paper proposes a deep learning multimodal fusion prognosis prediction model based on recurrent nasopharyngeal carcinoma PET/CT images, which significantly improves the prognosis prediction effect of recurrent nasopharyngeal carcinoma. Specifically, this paper uses the EfficientNet model and the DeepSurv loss function as the basis to construct a multimodal deep learning prognosis prediction model that fuses PET images and CT images. The model learns a set of deep learning features related to prognosis end-to-end with overall survival as the target and then models the Cox survival regression model. Experiments show that based on the prediction values of the multimodal fusion model, patients can be divided into low-risk and high-risk groups, with a 28.5% gain in 3-year survival rate (the difference in 3-year survival rate between low-risk and high-risk groups) compared to conventional clinical models (19.2%), which may enable these patients to choose more suitable treatment methods and improve their prognosis. This paper shows that deep learning can mine tumor heterogeneity information that clinical features do not possess from tumor imaging data, and may assist in the treatment decision-making of recurrent nasopharyngeal carcinoma.

(2) In response to the clinical problem of difficulty in early screening of nasopharyngeal carcinoma and the technical problem of privacy protection of multi-center nasopharyngeal carcinoma early screening data, this paper constructs a federated learning training framework for nasopharyngeal carcinoma early screening and realizes the training and application of a multi-center nasopharyngeal carcinoma early screening model that meets data security specifications. Specifically, based on the federated averaging strategy, this paper uses personalized layer federated strategy and federated-finetuning strategy to optimize federated learning training for the problem of non-independent and identically distributed multi-center data, enhancing the privacy of federated learning in a single center from the model and training perspectives. Compared with the locally trained models, more accurate early screening is achieved. Experiments show that due to the impact of non-independent and identically distributed data, the federated averaging strategy can only achieve a classification AUC of 0.8859, which is lower than the local learning of 0.8926. After optimization, the classification AUC of federated learning is0.8962, close to the centralized learning of 0.9023. This paper shows that federated learning can use multi-center nasopharyngeal carcinoma early screening data under the constraints of data security and privacy protection, achieving better early screening results than locally trained models.

(3) In response to the severe non-independent and identically distributed nature of the multi-center nasopharyngeal carcinoma early screening dataset and the poor generalization performance of the federated learning early screening model on the multi-center dataset, this paper proposes an adaptive data resampling algorithm based on model update distance, realizing the training and application of the federated learning model on the severely non-independent and identically distributed data. Specifically, based on the previous research, this paper calculates the resampling coefficient of each center by reducing the difference in the update distance of each center, with the global average value of the update distance as the target, and proportionally resamples each center’s dataset, so that the model can learn more small, difficult-to-learn data, thereby improving the generalization performance of the model on the multi-center dataset. The experiments show that the proposed algorithm significantly improves the classification AUC of the federated learning model on the multi-center nasopharyngeal carcinoma early screening dataset.

In summary, this article focuses on the issues of prognosis prediction and early screening of nasopharyngeal carcinoma. It proposes an image multimodal fusion prognosis prediction model and a distance-adaptive resampling algorithm based on model updating for local learning and federated learning, respectively. This achieves nasopharyngeal cancer assisted diagnosis and treatment based on local learning and federated learning, providing an effective tool to improve patient survival. The article was published as the first author (including co-first author) in the clinical mainstream journals European Journal of Nuclear Medicine and Molecular Imaging and Therapeutic Advances in Medical Oncology, and a patent was applied for and two software copyrights were registered.