脑部放射外科治疗中的微小病灶及器官智能分割技术研究

Stereotactic radiosurgery is one of the most important methods to treat brain functional abnormalities and small tumors. The key step of the treatment is the accurate delineation of the therapeutic targets. Essentially speaking, it’s lesion and organ segmentation based on medical images. Because the corresponding lesions and organs are always small and the treatment is sensitive with the contouring time, it’s difficult to segment manually. We proposed an automatic target delineation model to facilitate treatment planning of brain radiosurgery. During the following research, we found its accuracy and sensitivity outperform other similar methods, but it’s easy to over segmentation. After well study, we propose to work on the intelligent segmentation technology of small lesions and organs in brain radiosurgery based on multi-modality Magnetic resonance images. We intend to conduct the research work from the following aspects.(1)The multiscale features will be fused to perform the joint prediction to improve the robustness of lesion segmentation.(2)Based on capsule net, the quantitative representation of the relationship among organs will be presented. And the corresponding model will perform the accurate segmentation of small organs.(3)The mapping between the multimodality space and the full modality space will be constructed. It will compensate the missing modality samples and ensure the workflow works well. Through the research of the project, we hope to provide the critical technology supporting tools for brain radiosurgery. Meanwhile, it provides available solutions to other small target segmentation missions.

立体定向放射外科是治疗脑部功能性异常和微小肿瘤最为重要的手段之一，其成功实施的关键在于治疗目标的高精度勾画，本质是在医学影像上进行病灶和器官分割。由于相关病灶和器官体积微小、对勾画耗时敏感，手工分割十分困难。前期，我们构造了自动靶区勾画模型来辅助放射外科制定治疗计划，后续研究中发现其精度和敏感性在同类方法中表现尚佳，但易“错分割”。经过充分的预研实验，本项目提出基于多模态磁共振图像，针对脑部放射外科治疗中微小病灶和器官的智能分割技术进行研究，拟从以下几方面开展工作：融合多尺度信息进行联合预测，提高病灶分割的准确性；基于胶囊网络对器官间相互关系进行量化描述并构造分割模型，实现微小器官高精度分割；构造多模态空间到完整模态空间的映射，通过该映射对缺失模态样本进行信息补偿，保障多模态分割顺利实施。通过本项目的实施，将为脑部放射外科治疗提供关键技术支持，同时为相关领域的小目标分割任务提供解决思路。

立体定向放射外科是治疗脑部功能性异常和微小肿瘤最为重要的手段之一，其成功实施的关键在于治疗目标的高精度勾画，本质是在医学影像上进行病灶和器官分割。由于相关病灶和器官体积微小、对勾画耗时敏感，手工分割十分困难。前期，我们构造了自动靶区勾画模型来辅助放射外科制定治疗计划，后续研究中发现其精度和敏感性在同类方法中表现尚佳，但易“错分割”。经过充分的预研实验，本项目提出基于多模态磁共振图像，针对脑部放射外科治疗中微小病灶和器官的智能分割技术进行研究，主要开展了以下几方面的工作：构造了一种新的coarse-to-fine的分割框架DeSeg，通过将对象级的目标检测结合到像素级的图像分割任务中，融合多尺度信息提高微小病灶分割的精度和准确性；提出中心点引导SSD检测模型对器官间相互关系进行量化描述，提出能够根据图像特征自由选择信道扩张率的多尺度卷积算子并构造分割模型，实现微小器官高精度分割；构造多模态空间到本征特征空间的映射，通过该映射对缺失模态样本进行信息补偿，保障多模态分割顺利实施。结果显示，本课题所提出的病灶及器官分割算法的性能较现有算法有明显提高，同时，本课题所提出的多尺度卷积算子在医学影像分析相关的其他任务中也表现良好。课题研究成果已经发表了多篇学术论文，多次参加国内外学术会议进行宣传。通过本项目的实施，将为脑部放射外科治疗提供关键技术支持，同时为相关领域的小目标分割任务及医学影像分析的相关任务提供解决思路。