中国科学院自动化研究所机构知识库

近20年来，放射治疗技术发展迅速。与三维适形放射治疗(Three dimensional conformal radiation therapy, 3D-CRT)相比，调强放射治疗(Intensity-modulated radiation therapy, IMRT)和容积旋转调强治疗(Volume modulated arc therapy, VMAT)可更好地定位肿瘤靶标和保护危及器官，但IMRT和VMAT的缺点是计划设计更为复杂。为保证计划的安全性，治疗前需要开展一系列的质量保证(Quality assurance, QA)和质量控制(Quality control, QC)措施。随着人工智能技术在医疗领域的逐步推广，研究机构已经开始结合机器学习算法，致力于将传统的QA和QC流程自动化完成。其中，在确定放疗计划是否可以足够准确地应用于患者治疗时，选择合适的伽玛标准和干预限值至关重要。AAPM TG 218报告建议在3％/2mm伽玛标准下使用95％和90％分别作为 γ 通过率的容差限值和干预限值。因此，机器学习模型的最重要功能是在剂量验证之前找到可能无法通过容差限值/干预限值的放疗计划，以便进行更正。然而目前，在不同伽马标准的容差限值和干预限值下，机器学习模型的分类准确性尚未得到充分研究和验证。现有的基于机器学习的单中心放疗预测模型精度也有待进一步提升。 此外，在多中心大数据场景下，影响机器学习模型预测精度的因素以及如何提高多中心交互验证精度也有待深入研究。

针对上述问题，本论文提出了两种新的基于深度学习的预测分类算法，改进了已有的基于传统机器学习算法的预测模型，用于放疗计划患者剂量验证(Patient-specific quality assurance, PSQA)的结果评估，并在多中心较大样本中开展了交叉验证，提高了调强放疗患者剂量验证的精度。论文的主要创新性工作归纳如下：

1.提出了一种在治疗前进行患者 VMAT 剂量验证结果预测的机器学习框架(Possion+LASSO, PL)，该 PL 模型可以准确地预测大多数 VMAT 计划在 3％/3mm 和 3％/2 mm 的 γ 通过率。在 3％/3 mm、3％/2 mm 和 2％/2 mm 三个标准下，预测误差低至 1.81％, 2.39％和 4.18％。本研究采用了国际同类分析中的较大样本———共计 303 个 VMAT 计划 (包含了 176 个妇科肿瘤计划和 127 个头颈部肿瘤计划)。研究预测了两种肿瘤治疗部位混合的 VMAT 计划的 γ 通过率，并在不同标准下的进行分类研究，测试了灵敏度(Sensitivity) 和特异性 (Specificity)。结果证实，对于 VMAT 患者剂量验证而言，PL 模型是一种有效的辅助预测工具, 有望减少临床质量保证工作中医师和物理师的工作量，具有一定的临床应用价值。

2.提出了一种基于随机森林 (Random forest, RF) 的分类模型框架。该模型首先通过主成分分析法 (Principal component analysis, PCA) 对数据进行降维和特征提取，减少过拟合。之后使用随机降采样方法生成 1000 个不同的随机树，并最终利用集成学习的方法对生成随机树的决策进行集成，增加结果稳定性。实验结果表明，所提出的 RF 分类算法可以为 VMAT 计划患者剂量验证提供一种有效评估工具 (合格/不合格)，取得接近 100％的 VMAT 剂量验证灵敏度，并在完全独立的临床验证集上也取得了良好的预测结果，有望减少临床实际工作中患者剂量验证所需的工作量。

3.开发了一种基于深度学习的调强放疗患者剂量验证的分类和回归模型ACLR (Autoencoder based CLassification-Regression deep learning model)，该方法使用分类与回归相结合的框架，实现了在同一模型中预测 VMAT 计划在 3 种不同标准下的 伽马通过率以及是否合格 (分类)。该模型在技术验证和临床验证方面比之已有的Possion Lasso模型具有更高的准确性，具体而言技术验证和临床验证的准确性分别提高了 12-17％和6-17％。且ACLR 模型在 3％/ 3mm 伽玛标准下达到了接近 100％的灵敏度和 83％的特异性。

4.提出了一种新的基于 VAE(Variational Autoencoder) 和 CycleGAN 的深度学习模型 MSVCGAN (Multi-Sites Variational Autoencoders CycleGAN)。该算法的核心思想在于将不同中心 (机构) 的数据经过 VAE 编码器后获得的隐变量视为多中心数据间的通用表征形式，进而利用 CycleGAN 对 2 个不同风格间的数据进行迁移与转换，并进一步扩展到多个中心数据间的互相转换。基于该算法，我们开展了多中心大样本的放疗预测和质量保证研究，整合了 7 个中心共1835个VMAT 计划,是截止目前最大的多中心数据。研究证实了提出的MSVCGAN模型的临床可行性，改进了前述 ACLR 模型由于多中心差异导致的分类性能下降问题，可以准确地预测多中心场景下的患者剂量验证精度,验证了 MSVCGAN 模型在多中心大数据应用场景中的有效性。通过对 7 个中心的数据进行多种类别的分组和交叉独立验证，本文考察了加速器、剂量验证设备、放疗计划系统和不同采集中心对预测结果的影响。发现加速器、放疗计划系统的差异对预测和分类结果影响较大，而剂量验证设备的影响相对较小。因此，后续工作中将考虑对来自不同加速器、放疗计划系统的数据进行统一建模，增进所提预测模型的适用范围。

综上，将机器学习技术应用于放化疗剂量验证已呈现方兴未艾之势，本文围绕放疗大数据的患者剂量验证开展了一系列方法学研究，提出或改进了四种预测/分类算法，改善了在不同伽马标准下的PSQA预测精度,并尝试将模型推广到多中心大数据的应用场景。该研究有望减轻物理师临床实践中患者剂量验证工作的复杂性和工作量，具有重要的临床应用价值，也有待将来进一步的完善和补充。

In the past 20 years, radiotherapy technology has developed rapidly. Compared with Three dimensional conformal radiation therapy (3D-CRT), Intensity-modulated radiation therapy (IMRT) and Volume modulated arc therapy (VMAT) can be better To locate the tumor target and protect the organ at risk, but the disadvantage of IMRT and VMAT are that the plan design is more complicated. In order to ensure the safety of the plan, a series of quality assurance (QA) and quality control (Quality control, QC) measures need to be carried out before treatment. With the gradual promotion of artificial intelligence technology in the medical field, research institutions have begun to combine machine learning algorithms and are committed to automating the traditional QA and QC processes. Among them, in determining whether the radiotherapy plan can be applied accurately enough to the patient's treatment, it is very important to select the appropriate gamma criteria and action limit. The AAPM TG 218 report recommends using 95％ and 90％ as the tolerance  limit and action limit of the γ pass rate under the 3％/2mm gamma criteria, respectively. Therefore, the most important function of the machine learning model is to find a radiotherapy plan that may not pass the tolerance limit/intervention limit before patient treatment in order to make corrections. However, at present, under the tolerance limits and intervention limits of different gamma criteria, the classification accuracy of machine learning models has not been fully studied and verified. The accuracy of the existing single-center radiotherapy prediction model based on machine learning needs to be further improved. In addition, in the multi-center big data scenario, the factors that affect the prediction accuracy of the machine learning model and how to improve the accuracy of multi-center interactive verification are also to be studied in depth.

In response to the above problems, this paper proposes two new predictive classification algorithms based on deep learning, which improve the existing predictive models based on traditional machine learning algorithms for patient-specific quality assurance (PSQA) in radiotherapy plans. The results were evaluated, and cross-validation was carried out in a large sample of multiple centers, which improved the accuracy of patient treatment for IMRT patients. The main innovative work of the thesis is summarized as follows:

1. Proposes a machine learning framework (Possion+LASSO, PL) for predicting patient VMAT dose verification results before treatment. This PL model can accurately predict most VMAT plans at 3％/3mm and 3％/2 mm The γ pass rate. Under the three criteria of 3％/3 mm, 3％/2 mm and 2％/2 mm, the prediction errors are as low as 1.81％, 2.39％ and 4.18％. This study used a larger sample from the international similar analysis - a total of 303 VMAT plans (including 176 gynecological cancer plans and 127 head and neck cancer plans). The 255 plans were formed into the training set, and the leave-one-out cross-validation was carried out, and the unbiased verification test was performed on the other 48 clinical plan validation sets. The study predicted the γ pass rate of the VMAT plan mixed with two tumor treatment sites, and performed classification studies under different criteria, testing sensitivity (Sensitivity) and specificity (Specificity). The results confirmed that for VMAT patient dose verification, the PL model is an effective auxiliary prediction tool, which is expected to reduce the workload of physicians and physicists in clinical QA work, and has certain clinical application value.

2. Propose a classification model framework based on Random forest (RF). The model first uses Principal Component Analysis (PCA) to reduce dimensionality and feature extraction of the data to reduce over-fitting. After that, the random down-sampling method is used to generate 1000 different random trees, and finally the ensemble learning method is used to integrate the decision to generate the random tree to increase the stability of the result. Experimental results show that the proposed RF classification algorithm can provide an effective assessment tool (pass/fail) for VMAT plan patient dose verification, achieve a VMAT dose verification sensitivity of close to 100％, and also on a completely independent clinical verification set. Good prediction results have been achieved, which is expected to reduce the workload required for patient dose verification in clinical practice.

3. Developed a deep learning-based classification and regression model ACLR (Autoencoder based CLassification-Regression deep learning model) for dose verification of intensity-modulated radiotherapy patients. This method uses a framework that combines classification and regression to achieve in the same model Predict the gamma pass rate and eligibility (classification) of the VMAT plan under 3 different criteria. Compared with existing models, the ACLR method proposed in this paper divides the training set into different subsets according to the size of the gamma pass rate, and conducts training and verification on all data (576 VMAT plans) and different subsets; Among them, 426 plans are used as training data sets, and 150 plans are used as independent clinical test sets. Use ACLR and ten-fold cross-validation on the training data set to select the optimal hyperparameters and optimal model, and verify on the clinical test set. This model has higher accuracy than the existing Possion Lasso model in terms of technical verification (TV) and clinical verification (CV). Specifically, the accuracy of TV and CV are improved by 12-17％ and 6-17％, respectively. And the ACLR model achieves a sensitivity of close to 100％ and a specificity of 83％ under the 3％/3mm gamma criteria.

4. Propose a new deep learning model MSVCGAN (Multi-Sites Variational Autoencoders CycleGAN) based on VAE (Variational Autoencoder) and CycleGAN. The core idea of ​​the algorithm is to treat the hidden variables obtained from data from different centers (organizations) through the VAE encoder as a universal representation form among multi-center data, and then use CycleGAN to migrate and convert data between two different styles. And further expand to the mutual conversion between multiple center data. Based on this algorithm, we have carried out a multi-center large sample of radiotherapy prediction and PSQA(Patient-specific quality assurance) research, integrating a total of 1835 VMAT plans from 7 centers, which is the largest multi-center data so far. The study confirmed the clinical feasibility of the proposed MSVCGAN model, improved the aforementioned ACLR model's classification performance degradation caused by multi-center differences, and can accurately predict the accuracy of PSQA  in multi-center scenarios, and its performance is excellent The prediction result of each single center is at least equivalent to its performance, which verifies the effectiveness of the MSVCGAN model in multi-center big data application scenarios. By grouping and cross-independently verifying data from 7 centers, this article examines the impact of accelerators, dose verification equipment, radiotherapy planning systems (TPS), and different collection centers on the PSQA prediction results. It is found that the difference between accelerator and TPS has a greater impact on prediction and classification results, while the impact of dose verification equipment is relatively small. Therefore, in the follow-up work, unified modeling of data from different accelerators and TPS will be considered to improve the scope of application of the proposed prediction model.

In summary, the application of machine learning technology to the dose verification of radiotherapy and chemotherapy has been in the ascendant. This article has carried out a series of methodological studies on the patient dose verification of radiotherapy big data, proposed or improved four prediction/classification algorithms, and improved the PSQA prediction accuracy under different gamma criteria, and try to extend the model to multi-center big data application scenarios. This research is expected to reduce the complexity and workload of patient dose verification in the clinical practice of physicists. It has important clinical application value and needs to be further improved and supplemented in the future.