基于注意力机制对抗模型的CT-MRI跨模态肝脏肿瘤靶区勾画方法研究

Morbidity of hepatocellular carcinoma in China is increasing in recent years. Stereotactic body radiation therapy (SBRT) is emerging as a new significant non-invasive approach for hepatocellular carcinoma(HCC), Accurate target delineation of tumor is crucial step to the success of SBRT treatment. In order to reduce interobserver variability in target delineation caused by the partial volume effect (PVE) and improve accuracy.of tumor boundary location, this project focused on automatic target delineation for HCC based on cross-modal synthesis technology with magnetic resonance imaging (MRI) and computed tomography (CT). 1) Develop an algorithm with learning data augmentation and pixel-level label generated to conquer insufficient and imbalance of high quality large training data set. 2) Study theory and mechanism of domain shift between CT and MRI to develop a targetdelineation model based on attention-aware generative adversarial networks with synergistic fusion of adaptations from both image and feature perspectives as well as tumor shape variability context information in the CycleGAN towards to automatic tumor location precisely. 3) A unified 3D tumor segmentation benchmark was established to reuduce interobserver variability and verify the algorithm performance and clinical usability. The outcome of this project will provide valuable tools for evaluation of the uncertainties risk in local control rates and radiation dosage toxicity.

立体定位放射治疗(SBRT)近年来成为肝癌的重要治疗新手段，对肿瘤靶区勾画是SBRT治疗的关键步骤。针对肿瘤靶区边界受部分容积效应导致勾画的不一致性和跨模态位移引起定位不准确性的问题，本项目基于软组织高对比度成像的MRI增强肿瘤CT边界定位优势，围绕MRI-CT跨模态合成技术，开展靶区全自动勾画的关键技术研究：1）基于对抗网络的数据增强和标注自动生成模型，研究解决高质量训练数据不足和不平衡问题；2）研究CT-MRI跨模态位移原理及方法，实现基于注意力机制对抗分割模型，提出了一种融合图像域外观、特征域适配及形状域自适应的对抗学习模型，实现肿瘤靶区边界准确勾画；最后3）建立一套统一的肿瘤分割算法benchmark测试框架，有效降低不同专家之间勾画的显著差异性问题。研究成果将为肝脏肿瘤放疗的局部控制率及毒副作用等不确定性风险的定量分析提供新的技术手段。

随着成像与定位技术的快速发展，放射治疗近年来成为癌症的重要治疗方式，如何实现精准治疗是当前的研究热点。对肿瘤靶区勾画是放疗的关键步骤，临床医生需要在计划CT中勾画出靶区和危及器官的轮廓，这一步骤极为耗时耗力。肿瘤靶区的精准勾画是可保障放疗计划系统能够有针对性地将剂量准确投射到病灶区域，但是现有的靶区勾画算法在临床应用上还存在肿瘤靶区勾画精度低，跨模态域位移导致肿瘤边界勾画不准确，以及缺乏统一的benchmark方法评价算法性能等技术瓶颈问题。本研究由以下几个方面展开：（1）研究医学影像训练数据增强和标签自动生成，以获得高质量标注的肿瘤影像数据；（2）基于注意力机制进行CT-MRI跨模态融合分割的模型理论与实验研究；（3）在此基础上还将研究延伸拓展至数字病理图像智能分析；（4）通过肿瘤影像benchmark数据平台开发，建立有效且客观的评价方法，从而真正意义实现靶区自动勾画方法在临床上的推广应用。本项目累计发表文章11篇，申请专利3项，软件著作权1项；培养博士、硕士研究生5人；参加和组织学术交流4次。