肿瘤PET/CT/MRI纹理谱深度学习分析建模与智能放疗应用

在我国每年新增的400多万癌症患者中，70%以上的患者都需要进行放射治疗。肿瘤放疗靶区和危及器官的高精度勾画是智能放疗的前提，但目前还不能满足临床放疗精度要求。PET、CT、MRI影像纹理特征可提高勾画精度，但海量、复杂的肿瘤纹理谱难有效建模。本项目拟在天河一号超算平台上开展①肿瘤PET－CT－MRI靶区和危及器官勾画深度学习方法研究，提出联合肿瘤PET、CT、MRI影像纹理特征谱、集成CNN、SAE与PyraMiD-LSTM深度神经网络、构建肿瘤放疗靶区和危及器官检测、分类和识别混合深度神经网络DeepIRT，提出DeepIRT深度学习方法； ② PET、CT、MRI、CBCT图像高精度配准新方法研究；③肿瘤放疗靶区和危及器官高精度智能勾画、验证、评估临床研究。提出肿瘤放疗靶区和危及器官高精度勾画新理论和新方法，为PET、CT、MRI影像引导智能放疗提供科学支撑，具有重要临床和科学意义。

Each year, the estimated number of new cancer cases in China is over 4 million and 70% of them need radiotherapy. The high-precision delineation of the gross target volumes (GTV) and organs at risk (OAR) is the premise of the successful implementation of image-guided intelligent radiotherapy. However, the current methods for the automatic delineation of GTV and OAR in PET, CT and MRI images can’t meet the accuracy of clinical radiotherapy. Previous studies have indicated that, the texture features of PET, CT, MRI images of tumors can improve the delineation accuracy of GTV and OAR, but it is difficult to effectively identify and model the massive, complex texture spectrum of tumors..We plan to perform our project based on the Tianhe 1 Supercomputing Platform at National Supercomputing Center at Changsha. To high precisely delineate GTV and OAR，and to provide scientific support for intelligent radiotherapy, ① we propose to analyze and model the texture spectrum of PET, CT, MRI image of tumors by deep learning, and to design a hybrid deep neural network framework, namely DeepIRT, by combining CNN、SAE with PyraMiD-LSTM, and to learn the latent features of GTV and OAR by DeepIRT, ② to use new high-precision deformable elastic methods for the registration of PET, CT, MRI, CBCT images, ③ and new deep-learning based methods for the delineation of GTV and OAR in PET, CT, MRI images by combining clinical prior with DeepIRT, and to assess and verify the high-precision smart deep-learning based delineation methods by clinical research.