基于深度学习的胶质瘤IDH基因分型预测  

Prediction of Glioma IDH Genotyping Based on Deep Learning

作　　者：乔宝宝 彭博 戴亚康[2] 庞春颖[1] 李佳 QIAO Baobao;PENG Bo;DAI Yakang;PANG Chunying;LI Jia(School of Life Science and Technology,Changchun University of Science and Technology,Changchun 130022;Suzhou Institute of Biomedical Engineering Technology of Chinese Academy of Sciences,Medical Imaging Technology Research Laboratory,Suzhou 215163;China-Japan Union Hospital of Jilin University,Ward 3 of Neurology Department,Changchun 130031)

机构地区：[1]长春理工大学生命科学技术学院,长春130022 [2]中国科学院苏州生物医学工程技术研究所医学影像技术研究室,苏州215163 [3]吉林大学第三医院神经内科三病区,长春130031

出　　处：《长春理工大学学报（自然科学版）》2024年第5期82-88,共7页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62301557);江苏省重点研发计划项目(BE2022842,BE2022049-2);苏州市科技计划项目(SS202065)。

摘　　要：异柠檬酸脱氢酶(Isocitrate Dehydrogenase,IDH)基因突变是神经胶质瘤诊断与预后的重要生物指标,但当前临床检测该指标依赖昂贵的基因测序技术。针对以上情况,提出了基于深度学习的胶质瘤IDH基因自动分型方法。首先,该方法结合卷积神经网络与Transformer,利用已有胶质瘤IDH基因分型标签及其胶质瘤磁共振图像数据,训练初始分类网络;其次,进一步采用半监督学习方法,结合有、无IDH基因标签数据及其磁共振图像数据继续训练模型。实验结果表明,本方法提升了7%的分类精度,超越了其他经典算法和最新算法,为胶质瘤患者诊疗提供有力参考。The mutation of isocitrate dehydrogenase(IDH)gene is an important biomarker for the diagnosis and prognosis of gliomas,but the current clinical detection of this biomarker relies on expensive gene sequencing technology.To address this issue,an automatic typing method for glioma IDH gene based on deep learning algorithms is proposed.Firstly,the proposed method combines convolutional neural network and Transformer to train an initial classification network using the existing glioma IDH genotyping labels and glioma magnetic resonance image data.The semi-supervised learning method is used to train the model with and without IDH gene labeled data and magnetic resonance image data.Experimental results show that the proposed method can improve the classification accuracy by 7%,surpassing other classical algorithms and the latest algorithms,providing a valuable reference for the diagnosis and treatment of glioma patients.

关 键 词：异柠檬酸脱氢酶 深度学习 半监督学习 磁共振图像 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]