智能化内放疗手术规划系统关键技术研究

Brachytherapy is widely accepted for its minimally invasive character, irradiation-field fixation and good therapeutic effect. However, two key technologies are important factors to improving safety, planning efficiency and accuracy, reducing side effects and reducing complications of brachytherapy. One is intelligent segmentation and diagnosis of lung tumor with high automation, stability, accuracy and high efficiency. Another is intelligent dose planning technology based on clinical guidelines and priori success plans. In view of this, intelligent segmentation, malignant tumor detection and intelligent dose planning are the main subjects in this research. This study focuses on the clustering method, establishment of feature bank, optimal matching algorithm, intelligent dose planning method and optimization algorithm for benign and malignant tumors. First of all, the automatic selection algorithm of initial clustering center in lung tumor segmentation process is studied. Combined with the theory of fuzzy logic analysis, a fuzzy clustering fusion analysis method based on machine learning is developed to solve the automatic segmentation mechanism of lung parenchyma. By improving the statistical learning mechanism of support vector machine (SVM), intelligent detection of potential lung tumors can be realized. Secondly, based on the prior treatment plan data, the intelligent heterogeneous database is constructed, the dose optimization learning curve under multi-constraint conditions is studied, and the dose optimization model is established. Finally, the reliability and feasibility of the method are verified by comparative test and questionnaire survey. All these studies will support the development of an artificial intelligence-based brachytherapy planning system for lung cancer.

近距离粒子内放疗手术以其微创性、照射野固定、治疗效果好得到临床的广泛认可。然而，高精度且高效率的肺肿瘤智能分割、诊断技术和剂量规划技术的解决将在最大程度上提高手术安全性、规划效率和精确度，减少副作用、降低并发症。鉴于此，本研究以肺部肿瘤的近距离粒子内放疗手术规划为主要研究对象。针对聚类分析方法、肿瘤的特征库建立，最佳匹配算法，智能内放疗剂量规划及优化算法开展研究。首先研究肺部肿瘤分割过程中初始聚类中心自动选取算法，结合模糊逻辑分析理论，形成机器学习下的模糊聚类融合分析方法，以此解决肺实质的自动分割机理，通过改善支持向量机的统计学习机制研究，实现潜在肺部肿瘤的智能检测。其次，基于先验的治疗计划数据，构建智能异构数据库，研究多约束条件下剂量优化学习曲线，建立剂量优化模型。最后，通过对比试验及问卷调查，验证方法的可靠性和可行性。形成一套肺部肿瘤近距离粒子治疗计划系统。

近距离粒子植入手术作为肺肿瘤治疗的重要医疗手段，近年来被广泛应用于临床治疗中。然而，术前肺肿瘤诊断、分割及剂量计划制定的精度与效率是影响此类手术安全性及临床效果的关键。鉴于此，本项目研究团队通过将大数据、人工智能与临床医学交叉融合，针对肺肿瘤术前智能检测及智能分割机理、智能化术前剂量规算法展开深入研究：.1）完成了面向肺肿瘤的智能分割方法的构建。通过建立训练数据库，融合聚类分析体制与模糊逻辑理论，克服初始聚类中心的选取敏感等问题，综合考虑肺实质的形态学特征，并针对不同组织特性的肺实质去噪机理展开研究，实现肺肿瘤的术前自动智能分割。.2）实现了一种面向肺肿瘤的智能检测新方法。建立先验良恶性肿瘤特征数据库，通过对良恶肿瘤的纹理特性、离心率等特征进行综合分析考量，设计基于残差神经网络的肺肿瘤智能检测体系，并结合挤压-激励操作，突破小样本、高维非线性情况的下的肺肿瘤检测难题，实现肺肿瘤的自动智能检测。.3）构建了一种基于先验病例数据库的智能剂量规划理论体系。完成基于先验离散治疗计划数据的智能数据库的建立，以此为基础，建立基于孪生神经网络的术前病例相似性剂量预测方法，并提出一种基于模拟退火算法的剂量优化算法，二者有机结合，实现术前自动智能的剂量规划。.4）建立了一体化肺肿瘤术前智能诊疗系统，并搭建实验平台，结合临床先验成功病例数据库，对其进行临床可行性与安全性的验证。实验结果表明，本项目完成的智能诊疗系统能够对患者进行有效的术前精准诊断和手术规划。