基于集成学习的PET/CT混合成像肺癌检测  

Detection of lung cancer using PET/CT hybrid imaging based on integrated learning

作　　者：张瑞 程超 沈琳琳[3] 左长京 ZHANG Rui;CHENG Chao;SHEN Linlin;ZUO Changjing(Department of Software,Shenzhen Institute of Information Technology,Shenzhen,Guangdong 518172,China;Department of Nuclear Medicine,Shanghai Changhai Hospital,Second Military Medical University,Shanghai,200433,China;College of Computer Science and Software Engineering,Shenzhen University,Shenzhen,Guangdong 518060,China)

机构地区：[1]深圳信息职业技术学院软件学院,深圳518172 [2]海军军医大学长海医院核医学科,上海200433 [3]深圳大学计算机与软件学院,深圳518060

出　　处：《南昌大学学报（理科版）》2022年第6期666-673,共8页Journal of Nanchang University(Natural Science)

基　　金：国家自然科学基金重大研究计划项目(91959108);国家自然科学基金青年科学基金项目(11701389);广东省自然科学基金项目(2018A030313382);深圳市科技创新委员会基础研究项目(JCYJ20170306095702695)。

摘　　要：提出了一种基于多模态多尺度的Mask R-CNN集成学习模型对PET/CT混合成像进行人工智能肺癌检测。首先,通过5个深度学习模型对肺癌候选区进行提取。5个深度学习模型通过对不同尺度及不同模态训练数据进行Mask R-CNN迁移学习生成。然后利用集成学习方法将5个Mask R-CNN模型进行加权投票,有效减少假阳性数量,最终实现肺癌确诊。实验数据包括69例肺癌患者及11例正常例,训练数据集包括1242个肺癌横断面;验证数据包括270个横断面,其中58个PET肺癌横断面和58个CT肺癌横断面,77个PET正常横断面和77个CT正常横断面。该方法的F-score、Precision和Recall为0.95、0.90和1,与单模型和现有方法相比,本文方法对于PET/CT混合成像的肺癌检测具有更强的有效性,可以为医生提供有意义的辅助诊断信息.The issue of artificial intelligent detection for lung cancer in PET/CT hybrid imaging was investigated in this paper by using an integrated learning model based on multi-scale and multi-modality Mask R-CNN.Firstly, the candidate of lung cancer was extracted through three deep learning models.These three models were generated by appropriately tuning the Mask R-CNN employing certain training data that consisted of images from three different scales and different modalities.Then these three models were integrated using integrated learning of weight voting strategy to diminish the false positive outcomes.69 patients with lung cancer and 11 normal cases were included in the experiment.1242 slices with lung cancer were utilized for three training data sets.270 axial slices, including 58 PET slices and 58 CT slices with lung tumor and 77 PET normal slices and 77 CT normal slices, were used for testing.The F-score, precision and recall of resemble learning were 0.94,0.93,and 0.94,respectively.Compared with single model of Mask R-CNN and model of major voting, model of integrated learning based on multi-modality and multi-scale Mask R-CNN performed best.The experimental results showed this method is effective for the detection of lung cancer with PET/CT image, and can provide meaningful auxiliary diagnostic information for doctors.

关 键 词：人工智能 肺癌 集成学习 深度学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]