毕业生知识库

        同步放化疗 (Concurrent Chemoradiation Therapy, CCRT) 是局部晚期食管癌 (Locally Advanced Esophageal Cancer, LAEC) 患者的标准治疗方案。然而，仍有超过一半的患者在接受 CCRT 后出现局部复发。局部复发是导致患者死亡最常见 的治疗失败模式，准确预测 LAEC 患者的局部复发风险有助于辅助临床制定个 性化诊疗方案。但是受肿瘤异质性影响，目前临床上无法对 LAEC 患者的局部复 发风险进行有效预测。因此，迫切需要开发一种泛化性能好的预后标志物，准确 识别潜在的局部复发高风险患者。

        本研究收集来自 12 家医院的 LAEC 患者 CCRT 前后的计算机断层扫描 (Computed Tomography, CT) 影像，利用深度学习方法预测患者的局部复发风险。 相较于基于预定义影像特征和临床因素构建的预后模型，深度学习模型 Siamese-LSTM取得了最优预测性能。对食管癌数据集的定量和定性分析均显示，修正数据集间存在的图像异质性，是进一步提升深度学习预后模型泛化性能的可靠方案。 因此，我们研发了一个结合域自适应模块和 Siamese-LSTM 模块的两阶段食管癌同步放化疗预后预测模型。实验结果表明两阶段模型在独立验证集上取得了更优 的预测性能。

       本文的主要工作和贡献概括如下：

       1. 构建了一个基于治疗前后 CT 影像的食管癌同步放化疗预后预测模型。

       食管肿瘤随着时间推移而动态变化，纳入多时间点影像可能可以捕获更为全 面的肿瘤变化信息，从而准确预测患者局部复发风险。因此，我们纳入了患者 CCRT 前后的 CT 影像并使用深度学习模型 Siamese-LSTM 来捕获多时间点信息 用于预测局部复发风险。对比基于预定义特征和临床因素构建的预后模型， Siamese-LSTM 取得了较优性能。基于 Siamese-LSTM 开发的影像标志物 SLS 可 以显著区分患者的局部复发风险。此外，对食管癌数据集的定性及定量分析显示， 修正训练集和外部验证集间的图像异质性，是进一步提升深度学习模型泛化性能 的一种可靠方案。

       2. 提出了一种基于对抗学习的用于减少医学图像异质性的域自适应模块。

       深度学习通常假设训练集 (源域) 和外部验证集 (目标域) 共享相同的数据分布。然而，不同中心、不同设备及参数收集的医学图像往往被不同数量和类型的噪声污染，导致域偏移现象。因此，我们提出了一种注意力引导的生成对抗网络 Attention-CycleGAN。具体而言，我们为 CycleGAN 加入自注意力机制和注意 力特征图损失，通过注意力引导模型更好地关注重点域差异区域，进一步提升域 迁移的质量。本研究在医学数据集上进行验证，结果证明，Attention-CycleGAN作为域自适应模块，可以有效缓解医学图像异质性，提高下游模型的泛化性能。 此外，消融实验和对比实验的结果证明了 Attention-CycleGAN 的创新表现。

       3. 构建了一个结合域自适应模块和 Siamese-LSTM 模块的两阶段食管癌同 步放化疗预后预测模型。

       针对食管癌数据集间存在的图像异质性和准确预测 LAEC 患者局部复发风 险的临床需求，我们将 Attention-CycleGAN 与 Siamese-LSTM 结合，提出了一个域自适应食管癌同步放化疗预后预测模型。首先使用 Attention-CycleGAN 进行源域到目标域的迁移，接着将域迁移后的样本输入 Siamese-LSTM 以预测患者局部复发风险。在 Attention-CycleGAN 的训练过程中，自动调节数据增强操作的应用概率，防止在小规模食管癌数据集上出现过拟合现象。实验结果证明，两阶段食管癌预后预测模型取得了最优预测性能。基于两阶段模型开发的影像标志物 DA-SLS 可以显著区分患者的局部复发风险。最后，一系列亚组分析证明了 DASLS 的鲁棒性，并证明了其在结合 CCRT 短期疗效的基础上，有望辅助临床筛选出可能从早期巩固治疗中获益的患者。

       综上所述，本文提出了一个结合域自适应模块和 Siamese-LSTM 模块的两阶 段模型用于准确预测患者的局部复发风险。通过结合临床因素和 DA-SLS，有助 于筛选可能受益于早期巩固治疗的患者，辅助制定个性化治疗方案。 

Concurrent chemoradiation therapy (CCRT) is the standard of care for patients with locally advanced esophageal cancer (LAEC). However, more than half of patients still experience local recurrence after CCRT. Local recurrence is the most common treatment failure leading to patient death. Accurate prediction of the risk of local recurrence in LAEC patients is helpful for clinicians in the formulation of individualized diagnosis and treatment plans. However, due to the impact of tumor heterogeneity, it is currently impossible to effectively predict the risk of local recurrence in LAEC patients in clinical practice. Therefore, there is an urgent need to develop a prognostic marker with good generalization performance to accurately identify potentially high-risk patients for local recurrence.

This study collected pre- and post-CCRT CT images of LAEC patients from 12 hospitals and used a deep learning method to predict the risk of local recurrence. Compared with prognostic models constructed based on predefined imaging features and clinical factors, the deep learning model named Siamese-LSTM achieved optimal predictive performance. Quantitative and qualitative analyses of the esophageal cancer dataset showed that correcting the image heterogeneity between datasets is a reliable approach to further improve the generalization performance of deep learning prognosis models. Therefore, a two-stage prognostic model combining a domain adaptive module and the Siamese-LSTM module was developed. Experimental results demonstrated that the two-stage model achieved better predictive performance on the independent validation set. The main work and contributions of this study are summarized as follows:

1. Constructed a prognostic prediction model of CCRT for esophageal cancer based on pre- and post-treatment CT images.

Esophageal tumors are dynamically change over time. Including multiple time points of CT images may capture more comprehensive tumor information, thus accurately predicting the risk of local recurrence of patients. Therefore, we included CT images of patient pre- and post- CCRT and used Siamese-LSTM to capture information from multiple time points to predict the risk of local recurrence. Compared to prognostic models constructed based on predefined features and clinical factors, Siamese-LSTM achieved better performance. The developed imaging marker called SLS can significantly discriminate the local recurrence risk of patients. In addition, qualitative and quantitative analysis of the esophageal cancer dataset showed that correcting the image heterogeneity between the training and external validation sets is a reliable solution to improve the generalization performance of the deep learning model.

2. Proposed a domain adaptation module based on adversarial learning for reducing heterogeneity in medical images.

Deep learning typically assumes that the training set (source domain) and external validation set (target domain) share the same data distribution. However, medical images collected from different centers, devices, and parameters are often polluted by different types and amounts of noise, leading to domain shift. Therefore, we propose an attention-guided generative adversarial network called Attention-CycleGAN, to address this issue. Specifically, we introduce self-attention mechanism and attention feature map loss to CycleGAN, which allows the model to better focus on the domain difference regions and further improve the quality of domain transfer. This study is validated on a medical dataset, and the results demonstrate that Attention-CycleGAN, as a domain adaptation module, can effectively alleviate medical image heterogeneity and improve the generalization performance. Moreover, the results of ablation and comparison experiments confirm the innovative performance of Attention-CycleGAN.

3. Constructed a two-stage prognostic prediction model of CCRT for esophageal cancer combining the domain adaptation module and the Siamese-LSTM module.

Given the image heterogeneity among esophageal cancer datasets and the clinical need for accurately predicting the risk of local recurrence in LAEC patients, we proposed a domain-adaptive prognostic prediction model by combining AttentionCycleGAN with Siamese-LSTM. Firstly, we used Attention-CycleGAN to transfer images from the source domain to the target domain, and then we input the domaintransferred images into Siamese-LSTM to predict the risk of local recurrence in patients. During the training process of Attention-CycleGAN, we automatically adjusted the application probability of data augmentation operations to prevent overfitting on the small-scale esophageal cancer dataset. Experimental results demonstrated that our twostage prognosis model achieved optimal predictive performance. the imaging biomarker DA-SLS which is developed by our two-stage model showed significant ability to differentiate the risk of local recurrence in patients. Finally, a series of subgroup analyses confirm the robustness of DA-SLS and demonstrate its potential to assist in the clinical selection of patients who may benefit from early consolidation therapy based on the short-term efficacy of CCRT.

To sum up, our paper presents a two-stage model combining a domain-adaptive module and Siamese-LSTM module to predict the risk of local recurrence in LAEC patients. By incorporating clinical factors and DA-SLS, this model can assist in identifying patients who are likely to benefit from early consolidation therapy, and thus facilitate the development of personalized treatment plans.