基于多模态深度学习的脉管浸润检测

DETECTION OF MICROVASULAR INVASION HEPATOCELLUAR CARCINOMA BASED ON MUTILMODAL DEEP LEARNING

李庚泽

11849330

硕士

计算机技术

程然

2020-05-30

2020-07-08

哈尔滨工业大学

深圳

肝细胞癌是所有疾病中发病率和死亡率排在最高的疾病之一，国家癌症中心 研究显示肝细胞癌发病率和死亡率在所有癌症中居前 5 位，严重地威胁到现代人 们的生命安全。目前针对肝细胞癌的治疗手段主要是手术切除患病肿瘤区域。肝 细胞癌手术后复发的可能性很高，根据调查，我国进行肝细胞癌手术后，五年内复 发的概率高达 60%~70%。脉管浸润对判断手术切除后肝细胞癌是否复发有重要的 参考意义。在当前治疗水平下对肝细胞癌患者是否发生脉管浸润进行检测，对后 期的手术操作也有重要的指导意义。目前进行脉管浸润检测主要是传统的生化指 标检测，随着医疗影像技术的进步，基于影像来判断肝细胞癌脉管浸润的研究方 法开始出现。近年来人工智能技术有了较好的发展，机器学习，深度学习等技术 在图像、语音、文本等领域内获得了巨大的成功。医疗方面，深度学习已经在肺癌 结节检测等方面有所应用。本文首次使用深度学习技术对肝细胞癌脉管浸润进行 自动化检测。 本文利用图像识别技术，结合生化数据和医学图像数据建立了一个基于多模 态数据的神经网络模型，用于对肝细胞癌脉管浸润进行预测。在此基础上开发了 一套基于 web 服务的脉管浸润检测系统，用于辅助医疗人员改进肝细胞癌治疗方 案。该模型主要包括四个部分，图像预处理模块，图像数据预测模块，生化数据预 测模块，集成投票模块。在图像预处模块中，主要对遮罩层进行切割，使得输入数 据更聚焦于所研究的病灶位置;图像预处理模块将医疗图像数据转换为灰度值数 据。然后再经过数据的标准化处理以便更好地适应网络的输入要求。图像数据预 测模块主要使用卷积神经网络抽取病灶特征并使用监督学习进行预测。为了解决 数据量较少的问题，本文将 3D 立体扫描图像进行断层方向的切割，以便获取到更 多二维图像。对肿瘤相近位置的切片单独训练模型，在图像特征阶段训练了独立 的四个卷积网络模型。集成投票模块对图像数据预测模块的识别结果和生化数据 预测模块的识别结果进行影响权重的自学习，最后使用集成投票阶段的输出结果 作为最终的预测结果。基于以上研究方法，本文构建了一个在线肝细胞癌脉管浸 润检测系统，面向医疗工作者，提供肝细胞癌脉管浸润自动化识别服务，系统基于 web 端，轻量化，计算能力可根据用户需求量弹性扩展。 本文使用的实验数据由南京鼓楼医院提供的肝细胞癌患者的腹部 CT 扫描图 像数据和生化数据组成。总计 262 例病例数据，每个病例数据均由专家绘制肿瘤病灶区域，所有 CT 扫描图像的断层切片总计 13116 张。另一部分是患者医学指标 检测出的生化数据。

Hepatocellular carcinoma is one of the highest morbidity and mortality among all diseases, which seriously threatens the lives of modern people. According to data released by the National Cancer Center in 2018, the incidence and mortality of liver cancer in men are ranked in the top five. The treatment of hepatocellular carcinoma is mainly surgical removal of the affected tumor area. The risk of recurrence after hepatocellular carcinoma surgery is very high. According to the survey, the probability of recurrence within five years after hepatocellular carcinoma surgery is as high as 60 % ~70 %. Vascular invasion is of great significance for the recovery of hepatocellular carcinoma after surgical resection . Hepatic cell carcinoma radical resection is performed to detect whether vascular invasion has occurred in patients with hepatocellular carcinoma, and it also has important guidance for subsequent surgical operations significance. Breakthroughs have been made in such fields as text. In terms of medical treatment, deep learning has been applied in medical aspects such as lung nodule detection. This article will use deep learning technology for the first time to automatically detect vascular invasion of hepatocellular carcinoma. In this paper, a multimodal vascular invasion detection model based on convolutional neural network is studied for abdominal CT tomography data images and biochemical data. It is used to assist doctors in diagnosis in order to provide guidance for hepatocellular carcinoma resection. The model mainly consists of four parts. The first part is image preprocessing, and the second part is image feature extraction module, the third part is biochemical data extraction module, the last part is integrated voting stage. In the image preprocessing stage, the mask layer is mainly cut, so that the input data is more focused on the location of the lesion under study. Finally, the data is standardized to better meet the input requirements of the network. The image feature module mainly uses a convolutional network to extract lesion features. In order to solve the problem of low data volume, this paper cuts the 3D stereoscopic scan images in the tomographic direction, and separately trains the slices of tumors at similar positions, and trains eight independent convolutional network models at the image feature stage. The integrated voting stage performs self-learning on the weight of the world results of the image feature module and the recognition results of the deepened data extraction module, and uses the output of the integrated voting stage as the final prediction result. Based on the above research results, this article builds an online vascular infiltration detection system that provides medical vascular infiltration automatic identification services to medical workers. The system is web-based, lightweight, and can be flexibly expanded according to demand. This article has performed experiments on the proposed algorithm. The experimental data includes two parts. One is the CT scan data of the abdomen of patients with hepatocellular tumors, which totals 262 case data. Each case data is drawn by experts to map the tumor lesion area, in which the vascular infiltration. A total of 13，116 tomographic slices were obtained from all CT scans. The other part is the biochemical data detected by the patient’s medical indicators.

深度学习 脉管浸润 肝细胞癌 图像处理 卷积神经网络

deeplearning vascularinfiltration hepatocellularcarcinoma imageprocess- ing convolutional neural network

中文

联合培养

南方科技大学

李庚泽. 基于多模态深度学习的脉管浸润检测[D]. 深圳. 哈尔滨工业大学,2020.