毕业生知识库

鼻咽癌是一种源于鼻咽粘膜上皮的恶性肿瘤。据报道，我国鼻咽癌患者占全球鼻咽癌患者的近一半。在初诊时，超过70%的鼻咽癌患者已发展为晚期，其更容易发生复发和转移，生存质量较差，且死亡率是早期鼻咽癌患者的3倍以上。精准的预后预测和个体化的治疗决策有助于改善晚期鼻咽癌患者的生存状况。

当前，临床上通常基于磁共振成像（magnetic resonance imaging，MRI）进行预后评估和治疗决策。MRI扫描可以全面、无创地反映鼻咽癌病灶及其周围组织的形态学和解剖学信息，而新兴的影像组学方法能够有效挖掘MRI影像中与预后和治疗效果相关的定量特征，为提升鼻咽癌的预后预测精度和治疗水平带来了新的机遇。然而，基于传统机器学习或者深度学习的影像组学方法，在鼻咽癌预后预测和治疗决策方面均存在技术缺陷。例如，在预后预测方面，现有研究并未充分利用生存时间信息，也未针对MRI影像特点设计与预后相关的特征提取网络；在治疗决策方面，常规的基于疗效预测的方法在不同亚组之间的稳定性较差，也无法预测患者接受特定治疗方式时的预后。针对上述问题，本文基于MRI影像开展了晚期鼻咽癌预后预测和治疗决策方法研究，辅助临床制定更优的个体化治疗方案。其中，治疗决策方法研究主要围绕单独同期放化疗（concurrent chemoradiotherapy，CCRT）和诱导化疗联合CCRT两者之间的选择问题。本文的主要工作和贡献如下：

1.   针对临床因子对晚期鼻咽癌的预后预测能力不足的临床问题，以及预后分析中的损失函数难以有效利用生存时间信息的技术问题，本文提出了一种基于时间自适应协调损失的晚期鼻咽癌多目标预后模型。相比于现有损失函数只能建模单尺度的生存时间信息，本文通过融合代表不同尺度信息的似然损失、改良的排序损失、校验损失，构建了时间自适应的协调损失，可以多角度地挖掘特征与预后之间的复杂关系。在此基础上，为了使该损失函数更好地适用于MRI影像分析，本文针对MRI影像固有的各向异性问题，设计了基于多示例学习的多目标预后模型，有效挖掘MRI影像中预后相关的特征。实验结果发现，相比于其他损失，基于该损失构建的模型在三个数据集上达到了最高的预后预测精度。此外，多目标预后模型产生了具有良好预后预测能力的影像组学标签，将其添加到临床模型中可以显著提高模型的晚期鼻咽癌预后预测精度（测试集上的C-index：0.788 vs. 0.625，学生t检验：p = 0.001）。

2.   针对诱导化疗的治疗反应在患者间存在差异的临床问题，以及当前的疗效预测方法稳定性差的技术问题，本文提出了一种基于比例风险回归的晚期鼻咽癌诱导化疗疗效预测模型。一方面，针对影像组学难以处理预后信息冗余的问题，本文设计了一种基于交互项的疗效敏感的特征选择方法，可以筛选出可重复性高、稳定性强且与疗效相关的关键特征。另一方面，为了融合疗效相关的关键特征，本文设计了一种基于特征变换的比例风险回归模型。相比于常规的疗效预测模型仅能建模近期治疗效果，提出的模型可以有效建模治疗方式间的远期生存差异。在前瞻性的测试集上，按照临床指南推荐的诱导化疗为晚期鼻咽癌患者带来了16.6%的生存获益，而根据本文模型的结果进行分层后，诱导化疗为晚期鼻咽癌患者带来的生存获益提升到了36.9%，优于其它疗效预测方法获取的生存获益。此外，通过融合本文模型的结果和临床风险因子，能够识别出适合CCRT的患者和需要接受更激进治疗方式的患者。

3.   针对晚期鼻咽癌缺乏最佳治疗方案选择标准的临床问题，以及现有疗效预测模型无法预测不同治疗方式下患者预后的技术问题，本文提出了一种基于多任务深度学习的晚期鼻咽癌治疗方案推荐模型。相比于现有模型无法建模预后与治疗疗效的关联，本文利用多任务学习方法解耦预后模型中治疗交互间的复杂关系，设计了一种能同时进行预后预测和疗效预测的治疗方案推荐模型。提出的模型可以预测患者在接受不同治疗方式时的预后，并据此推荐最有利的治疗方案。实验结果表明，在预后性能上，提出的模型优于对各治疗方式单独构建的模型。利用该决策推荐模型对患者进行分层，可以识别出5年生存率在91%以上的适合CCRT或诱导化疗联合CCRT的患者，而如果不进行分层，患者的5年生存率仅为84%，并且诱导化疗无法带来显著的生存获益。此外，提出的模型也识别出了需要寻找其他治疗方式的患者。

总体而言，围绕鼻咽癌的预后预测和治疗决策问题，本文通过创新影像组学特征提取和模型构建方法，针对临床需求和现有技术挑战，分别提出了多目标预后预测模型、诱导化疗疗效预测模型和治疗方案推荐模型，从MRI影像中提取与鼻咽癌预后和治疗疗效相关的关键影像特征，有望辅助治疗方案的选择。

Nasopharyngeal carcinoma (NPC) is an epithelial malignancy originating from the nasopharyngeal mucosa. It is reported that patients with NPC in China account for nearly half of the patients with NPC. Unfortunately, over 70% of patients with NPC have developed into advanced stage at the initial diagnosis, which is associated with a higher risk of recurrence and metastasis, lower quality of survival, and a mortality rate more than three times that of early-stage NPC patients. Accurate prognosis prediction and individualized treatment decisions can help improve the survival of patients with advanced NPC.

Currently, prognostic assessments and treatment decisions are usually based on magnetic resonance imaging (MRI) in clinical practice. As the preferred and non-invasive imaging tool, MRI scan can comprehensively reflect the morphological and anatomical information of nasopharyngeal cancer lesions and their surrounding tissues. Moreover, the emerging radiomics can effectively explore the quantitative features related to prognosis and treatment response from MRI images, thereby bringing new opportunities to optimize prognosis prediction and treatment decisions in NPC. Despite the potential of radiomics, traditional machine learning or deep learning-based radiomic methods have technical shortcomings in prognosis prediction and treatment decisions for NPC. For example, in terms of prognosis prediction, existing studies have not fully utilized survival time information or designed specific networks to effectively extract prognosis-related radiomic features for the characteristics of MRI images; in terms of treatment decisions, conventional methods for predicting treatment response have poor stability of predicted outcomes among different subgroups and cannot predict the prognosis of patients when they receive specific treatment regimens. To address the aforementioned issues, this thesis investigates the methods about prognosis prediction and treatment decisions for advanced NPC based on MRI images, aiming to assist clinicians in developing better individualized treatment plans. Specifically, the study about treatment decisions focuses on the choice between concurrent chemoradiotherapy (CCRT) alone and induction chemotherapy combined with CCRT. The main work and contributions of this thesis are as follows:

1. Aiming at the clinical problem that clinical factors have unsatisfactory prognostic ability in advanced NPC, and the technical problem that existing loss functions for prognosis prediction are difficult to effectively utilize survival time information, we proposed a multi-task prognostic model for NPC based on the time-adaptive coordinate loss. Unlike existing loss functions that can only learn survival time information in a single scale, the proposed time-adaptive coordinate loss fuses likelihood loss, modified ranking loss, and calibration loss to represent different scales of information. The proposed enables the prognostic model to learn the complex relationship between features and patients’ prognosis from multiple perspectives. Additionally, we designed a multi-task prognostic model based on multiple instance learning to address the inherent anisotropy problem of MRI images, allowing for effective extraction of prognostic information from MRI images. The experimental results found that the model trained using the proposed loss achieves the highest accuracy in prognosis prediction on three cohorts compared to other losses. In addition, the multi-task prognostic model outputs a radiomic signature with good prognostic predictive ability, and adding it to the clinical model improves the accuracy of prognosis prediction for advanced NPC (C-index on the test set: 0.788 vs. 0.625, Student's t-test: p = 0.001).

2. Aiming at the clinical problem of variation in treatment response to ICT between patients, and the technical problem of poor performance stability of current methods for treatment response, we proposed proposes a model based on proportional risk regression for predicting response to induction chemotherapy in advanced NPC. On the one hand, to address the problem that radiomics is difficult to deal with the redundancy of features, we designed a feature selection method based on interaction terms to identify key features that are reproducible, stable, and correlated with treatment response. On the other hand, based on feature transformation, we designed a proportional risk regression model to fuse the key features correlated with treatment response. Compared with commonly used response prediction models that can only model near-term treatment response, the proposed model can effectively model the long-term survival differences between treatment regimens. In the prospective test cohort, induction chemotherapy as recommended by clinical guidelines provided a survival benefit of 16.6% for patients with advanced NPC, while after risk stratification based on our model, the survival benefit of induction chemotherapy for patients with advanced NPC was increased to 36.9%, which was superior to that obtained by other response prediction models. Moreover, the integration of the model output with clinical risk factors allowed for the identification of patients suitable for CCRT and those who needed to undergo more aggressive treatment regimens.

3. Aiming at the clinical problem that there is lack of clinical criteria for selecting the best treatment regimen for advanced NPC, and the technical problem that existing methods for predicting treatment response cannot predict patients’ prognosis when they receive specific treatment regimens, we proposed a treatment recommendation method based on multi-task deep learning for advanced NPC. Compared with existing models that cannot model the association between prognosis and treatment response, we designed a treatment regimen recommendation model that can predict both prognosis and treatment response, by decoupling the complex relationships between treatment interactions in the prognosis model. The proposed model can predict patients’ prognosis when they receive different treatment regimens and recommend the favorable treatment regimen accordingly. Experimental results show that the proposed model outperformed existing models constructed separately in each treatment regimen group in terms of prognostic performance. After stratifying patients using the proposed model, we could identify patients suitable for CCRT or induction chemotherapy combined with CCRT, with a 5-year survival rate of over 91%; whereas without stratification, the 5-year survival rate of patients is only 84% and induction chemotherapy does not bring about a significant survival benefit. In addition, the proposed model identified patients who need to find other suitable treatment regimens.

In summary, this thesis addresses the challenges of prognosis prediction and treatment decision-making in NPC, and proposes three novel radiomic algorithms for feature extraction and model construction. These include a multi-task prognosis prediction model, a model for predicting response to induction chemotherapy, and a treatment recommendation method, all designed to effectively extract key radiomic features from MRI images that are related to the prognosis and treatment response of NPC, ultimately assisting in treatment selection for advanced NPC.