波谱学杂志 ›› 2018, Vol. 35 ›› Issue (4): 447-456.doi: 10.11938/cjmr20182658

• • 上一篇    下一篇

基于AI+MRI的影像诊断的样本增广与批量标注方法

汪红志1, 赵地2, 杨丽琴3, 夏天1, 周皛月4, 苗志英5   

  1. 1. 华东师范大学, 上海市磁共振重点实验室, 上海 200062;
    2. 中国科学院计算技术研究所, 北京 100190;
    3. 复旦大学附属华山医院, 上海 200040;
    4. 西门子医疗系统有限公司, 上海 201318;
    5. 上海理工大学 光电信息与计算机工程学院, 上海 200093
  • 收稿日期:2018-05-31 出版日期:2018-12-05 发布日期:2018-07-09
  • 通讯作者: 汪红志,Tel:13916346546,E-mail:hzwang@phy.ecnu.edu.cn;赵地,Tel:010-62601166,E-mail:zhaodi@escience.cn. E-mail:hzwang@phy.ecnu.edu.cn;zhaodi@escience.cn
  • 基金资助:
    浦江人才计划(17PJ432500).

An Approach for Training Data Enrichment and Batch Labeling in AI+MRI Aided Diagnosis

WANG Hong-zhi1, ZHAO Di2, YANG Li-qin3, XIA Tian1, ZHOU Xiao-yue4, MIAO Zhi-ying5   

  1. 1. Shanghai Key Laboratory of Magnetic Resonance, East China Normal University, Shanghai 200062, China;
    2. Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China;
    3. Huashan Hospital, Fudan University, Shanghai 200040, China;
    4. Siemens Healthineers, Shanghai 201318, China;
    5. School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
  • Received:2018-05-31 Online:2018-12-05 Published:2018-07-09

摘要: 训练样本是所有领域人工智能(AI)研发的关键因素.目前,基于人工智能+磁共振成像(AI+MRI)的影像诊断存在着训练样本的有效标注数量和类型无法满足研发需求的瓶颈问题.本文利用临床MRI设备对志愿者或阳性病例进行正常或重点病灶区的定量扫描,获取高分辨率各向同性的纵向弛豫时间(T1)、横向弛豫时间(T2)、质子密度(Pd)和表观扩散系数(ADC)等物理信息的多维数据矩阵,作为原始数据.开发虚拟MRI技术平台,对原始数据(相当于数字人体样本)进行虚拟扫描,实现不同序列不同参数下的多种类磁共振图像输出.选择感兴趣组织具有最好边界区分度的图像种类,经有经验的影像医生对其进行手动勾画并轨迹跟踪形成三维MASK标注矩阵,作为其他种类图像的图像勾画标注模板,从而实现低成本、高效率的MRI样本增广和批量标注.该平台以临床少量阳性病例作为输入,进行样本增广和标注,极大地减少AI对实际扫描样本的要求,降低了影像医生的精力和时间投入,极大地节省了成本,并输出了数量足够的磁共振图像,为基于AI+MRI的影像诊断研发提供低成本的训练数据解决方案.

关键词: 人工智能(AI), 磁共振成像(MRI), 样本增广, 批量标注, 影像辅助诊断

Abstract: Training data enrichment is a key factor in artificial intelligence (AI) technology development. At present, the bottleneck problem is that the quantity and type of labeled training data in valid samples are unable to meet the requirements of AI+MRI aided diagnosis. In this paper, an effective approach to solve the problem was presented. High resolution isotropic multi-dimensional data of regions of interests from patients or healthy volunteers were first acquired via a series of scanning on clinical MRI scanners, including quantitative T1, T2, proton density (Pd) and apparent diffusion coefficient (ADC) measurements. These data were then used as the ground truth, from which different types of images associated with different imaging sequences and parameters were obtained with a virtual MRI technology. The type of the images with the best boundary resolution were then selected manually by experienced doctors, on which three-dimensional mask matrix was obtained by manual contouring and labeling, serving as the template for other types of images. This enrichment method was developed as a software platform, which could provide sufficient quantity of image data from a small number of positive cases, thus meeting the data training enrichment requirement of AI+MRI diagnosis at low cost and with high efficiency.

Key words: artificial intelligence (AI), magnetic resonance imaging (MRI), training data enrichment, batch labeling, image aided diagnosis

中图分类号: