基于深度学习策略的超声多模态影像组学方法研究
周辉
2020-05-24
页数132
学位类型博士
中文摘要

临床上对于疾病的早期精准诊断对于患者病情的掌握,治疗方案的选择,预后的好坏起着关键的作用,有效的诊断方法往往能使病人得到最佳的治疗机会。目前临床上对于疾病的精准诊断主要是穿刺活检或者术中病理等方法。这些方法为各大临床指南所推荐,虽然其准确度高,但是却面临着有创、样本取样误差、伴随后遗症等难以消除的缺点。因此,临床上亟待发展无创精准的疾病诊断方法,辅助医生在临床上进行早期精准诊断。

随着医疗设备的更新,特别是成像技术的发展,目前利用影像技术对各大疾病进行无创的精准诊断越来越普遍,成为临床医生判断病情的主要手段之一。其中尤其以计算机断层扫描成像(Computed TomographyCT)、核磁共振成像(Magnetic Resonance Imaging, MRI)、X光片以及超声成像(UltrasoundUS)最为普遍。在某些浅表器官(如甲状腺和肝脏)的诊断方面,超声成像由于其价格便宜、无辐射以及操作简便快捷等优点而得到了广泛的应用。但是由于超声成像分辨率低、模态多样化和对操作者依赖性大,且无法定量描述病灶的影像学特征,其往往依赖于临床医生的经验和主观判断,因此对于疾病的精准判断带来了很大的挑战。

目前,一种基于人工智能的影像分析方法——影像组学,能够很好的克服上述超声影像方法所带来的挑战。影像组学主要是借助于人工智能方法和图像处理技术,从大量医学图像中高通量地提取并分析病灶的影像信息,能够揭示出人眼无法获取的高维度、较为丰富的信息,为临床提供辅助决策支持。特别是随着可用数据的增多和计算机能力的提升,以深度学习为代表的数据驱动型模型得到了更多的关注。目前,影像组学已经被广泛应用于CTMRI图像的分析当中,而基于超声图像的超声影像组学的分析尚处于早期发展阶段。因此,本文将利用深度学习的技术,对超声影像中两种不同模态(普通二维超声和弹性超声)进行分析,以达到精准诊断典型超声适用疾病(如肝脏和甲状腺)的目的,辅助临床医生实现早期精准诊断的目标。总的来说,本文将从以下几方面开展研究。

1)针对超声弹性成像弹性值标准不一的问题,本文以肝纤维化疾病为研究对象,研究使用卷积神经网络对肝纤维化进行精准分期。该方法利用数据数量多、质量优和来源广的特点,训练出针对该问题特异性高的自学习深度学习模型。相对于传统临床上的统计学分析方法和血清学分析方法来说,该方法能够定量的给出较为精准的分类结果,克服了以弹性测量值作为肝纤维化判断标准时的不一致问题。实验结果表明,该方法能够学习到数据中蕴藏的高通量信息,并得到精度高、鲁棒性强的判断结果。该方法在前瞻性多中心的数据样本中,取得了全面领先于临床方法的效果,特别是在严重肝纤维化和肝硬化这两个分类标准上,取得了逼近临床金标准的结果,其受试者工作特征曲线面积( Area Under Reciever Operating Characteristic Curve, AUC)分别可达 0.97 0.98

2)针对超声普通二维成像分辨率低、有效信息少的问题,本文以甲状腺结节为研究对象,提出多感兴趣区输入(multi region of interest)结合迁移学习模型,对甲状腺结节的良恶性进行判断。该方法通过多感兴趣区输入(multi-ROI)的模型结构,充分利用病灶区域周围以及其内部的信息。并结合前期工作的基础,将前期弹性超声的模型迁移到该模型中,使得该方法的结果得到进一步的提升。该方法不仅相对于临床医生具有更高的准确性,同时在不同仪器间也具有较好的鲁棒性。在该模型所基于的训练集和验证集上,AUC分别达到了0.960.95,证实了multi-ROI输入结合迁移学习的方法,对于甲状腺结节良恶性判断的有效性。

3)针对临床超声图像数据偏少以及来源方式的问题,本文以甲状腺结节为研究对象,研究使用在线迁移学习模型用于临床实践中,通过不断收集临床超声图像数据来提升模型的诊断效果,进一步达到辅助临床医生的作用。该方法在前期迁移学习的基础上,融入了在线学习的策略,使得模型更加稳定可靠,同时能够不断提升模型的效能。最终该方法的AUC值达到0.98,证实了该方法在提升了实用性的同时,也进一步提升了模型分类效果。

英文摘要

In clinic, early and accurate diagnosis of disease often plays a key role in the choice of treatment and the prognosis. A good diagnosis method can often provide the patient with the best treatment opportunity. At present, the accurate diagnosis is mainly the biopsy or intraoperative pathology. These methods are recommended by lots of clinical guidelines. Although they have high accuracy, they still are invasive and limited by sample errors, inter-observer variability, and various potential complications. Therefore, it is urgent to develop non-invasive and accurate disease diagnosis methods to assist doctors in early and accurate clinical diagnosis.

With the update of medical equipment, especially the development of imaging technology, it is common to use imaging technology for non-invasive accurate diagnosis of major diseases, which has become one of the main methods for clinicians to observe the disease. Among them, computed tomography (CT), magnetic resonance imaging (MRI), X-ray and ultrasound (US) are the most used technology. In the diagnosis of some superficial organs (such as thyroid and liver), ultrasound imaging has been widely used because of its advantages of low price, no radiation and simple operation. However, due to the low resolution of ultrasound imaging, the diversity of organs and the dependence of operator’s experience, it often depends on the experience and subjective judgment of clinicians, which brings great challenges to the accurate diagnosis of diseases.

At present, an image analysis method based on artificial intelligence, which is called radiomics, can solve the challenges brought by the above-mentioned ultrasound image methods. With the help of artificial intelligence method and image processing technology, we can extract and analyze high-throughput image features of lesions from a large number of medical images, which can reveal the rich information that can not be obtained by human eyes, and provide clinical decision support for clinical practice. With the increase of available data and the improvement of computing source, the data-driven model represented by deep learning has attracted more and more attention. Radiomics has been widely used in the analysis of CT and MRI, and the analysis of ultrasound image is still in the early stage of development. Therefore, this paper will use the deep learning technology to analyze two different modes of ultrasound image (ordinary two-dimensional ultrasound and elastography), so as to achieve the purpose of accurate diagnosis of typical ultrasound applicable diseases (such as liver and thyroid), and assist clinicians to achieve early accurate diagnosis. Specifically, this paper will carry out research from the following aspects.

(1) In order to solve the problem of different value standards of ultrasound elastography, this paper focus on the liver fibrosis disease, and using convolutional neural network to accurately stage liver fibrosis. This method, which is called deep learning radiomics of elastography (DLRT), makes full use of the characteristics of high quality data to train a self-learning deep learning model with high specificity for this problem. Compared with the traditional clinical statistical analysis and biomarkers analysis, this method can give more accurate classification results quantitatively. The experimental results show that the method can learn the high-throughput information contained in the image data, and get the results with high accuracy and strong robustness. This method showed similar diagnostic efficacy with the liver biopsy for assessing cirrhosis and advanced fibrosis which were significantly better than traditional clinical statistical analysis and biomarkers analysis.

(2) To solve the problem of low resolution and less effective information in two-dimensional ultrasound imaging, this paper proposes a multi region of interest (ROI) combined with transfer learning model, which is called deep learning radiomics of thyroid (DLRT), to diagnosis the benign and malignant thyroid nodules. This method makes full use of the features around and inside the thyroid nodules through the structure of multi ROIs. This method not only has higher accuracy than clinicians, but also has better robustness among different instruments. The area under the curve (AUC) of the training set and the validation set based on the model reached 0.96 and 0.95 respectively, which confirmed the validity of multi ROIs input combined with transfer learning in the diagnosis of benign and malignant thyroid nodules.

(3) In order to solve the problem of the amount of clinical ultrasound image data, this paper makes use of online transfer learning (OTL) model to further solve the problem. OTL counld achieve the goal of assisting clinicians through the continuous collection of clinical ultrasound image data to improve the diagnosis effect of the model. This method adopts the strategy of online learning and transfer learning, making the model more stable and reliable, and constantly improving the efficiency of the model.

关键词超声成像 深度学习 影像组学 无创诊断 人工智能
语种中文
七大方向——子方向分类医学影像处理与分析
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/38570
专题复杂系统管理与控制国家重点实验室_影像分析与机器视觉
推荐引用方式
GB/T 7714
周辉. 基于深度学习策略的超声多模态影像组学方法研究[D]. 中国科学院大学自动化研究所. 中国科学院大学,2020.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
博士论文-周辉.pdf(4006KB)学位论文 开放获取CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[周辉]的文章
百度学术
百度学术中相似的文章
[周辉]的文章
必应学术
必应学术中相似的文章
[周辉]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。