基于宏观与微观影像学特征融合的胶质瘤基因状态与分级预测研究  

Prediction of genetic status and grading in glioma based on fusion of macro-and micro-imaging features

作　　者：李臻 宋鹏斐 朱锐泽 江山 曹诗文 余锦华[2] 史之峰 LI Zhen;SONG Peng-fei;ZHU Rui-ze;JIANG Shan;CAO Shi-wen;YU Jin-hua;SHI Zhi-feng(Department of Neurosurgery,Huashan Hospital,Fudan University,Shanghai 201107,China;Grade 2023 of Department of Biomedical Engineering;Grade 2021 of Department of Biomedical Engineering;Department of Biomedical Engineering,School of Information Science and Technology,Fudan University,Shanghai200438,China)

机构地区：[1]复旦大学附属华山医院神经外科,上海201107 [2]复旦大学信息科学与工程学院生物医学工程系,上海200438

出　　处：《中国现代神经疾病杂志》2025年第3期165-174,共10页Chinese Journal of Contemporary Neurology and Neurosurgery

基　　金：上海市卫生健康委员会优秀项目(项目编号:20234Z0009);国家重点研发计划项目(项目编号:2022YFF1202804);国家自然科学基金资助项目(项目编号:82373018),国家自然科学基金资助项目(项目编号:82072020);上海市科委医学创新研究项目(项目编号:23Y11906200)。

摘　　要：目的 建立基于MRI与全视野数字切片(WSI)特征融合的双层特征蒸馏的多实例学习(DLFD-MIL)模型,实现对成人型弥漫性胶质瘤IDH1突变、1p/19q共缺失及世界卫生组织(WHO)分级的高精准性预测。方法 选择2021年1月至2024年6月复旦大学附属华山医院收治的212例成人型弥漫性胶质瘤患者及美国癌症基因组图谱计划42例成人型弥漫性胶质瘤病例,联合分析术前T2-FLAIR影像与术后WSI数据。构建DLFD-MIL模型,采用伪包生成策略解决WSI弱监督学习中的实例标签缺失问题,Concat融合方式实现多模态融合;绘制受试者工作特征曲线,以曲线下面积比较单模态与多模态特征融合对胶质瘤基因状态和WHO分级的预测效能。结果 在IDH1突变预测任务中,多模态特征融合模型的曲线下面积大于单模态WSI模型(Z=2.752,P=0.006)和单模态T2-FLAIR模型(Z=5.662,P=0.000);在1p/19q共缺失预测任务中,多模态特征融合模型的曲线下面积与单模态WSI模型(Z=-0.245,P=0.806)和单模态T2-FLAIR模型(Z=0.781,P=0.435)差异均无统计学意义;在WHO分级预测任务中,多模态特征融合模型的曲线下面积大于单模态T_(2)-FLAIR模型(Z=4.830,P=0.000),而与单模态WSI模型差异无统计学意义(Z=1.739,P=0.082)。结论 基于宏观与微观影像学特征融合模型可以提高胶质瘤IDH1基因分型和WHO分级的预测精度,为临床制定个性化治疗方案提供可靠的人工智能决策支持工具。Objective To develop a dual-layer feature distillation multiple instance learning(DLFDMIL)model integrating MRI and whole slide image(WSI)features for precise prediction of IDHI mutation,1p/19q codeletion,and World Health Organization(WHO)grading in adult-type diffuse gliomas.Methods A retrospective cohort of 212 adult-type diffuse gliomas patients from Huashan Hospital,Fudan University(January 2021 to June 2024)and 42 cases from The Cancer Genome Atlas(TCGA)were included.Preoperative T,-FLAIR and postoperative WSI data were jointly analyzed.The DLFD-MIL model addressed the lack of instance-level labels in weakly supervised WSI learning via a pseudo-bag generation strategy.Multimodal feature fusion was achieved through Concat.Diagnostic performance for molecular subtyping and WHO grading was evaluated by comparing area under the curve(AUC)of receiver operating characteristic(ROC)curve between single-mode(WSI or MRI)and multi-mode.Results In the IDHI mutation prediction task,AUC of the multi-mode feature fusion model surpassed single-mode WSI model(Z=2.752,P=0.006)and single-mode T,-FLAIR model(Z=5.662,P=0.000).In the 1p/19q codeletion prediction task,no statistically significant differences in AUC were observed between the multi-mode feature fusion model and either single-mode WSI model(Z=-0.245,P=0.806)or T_(2)-FLAIR model(Z=0.781,P=O.435).In the WHO grading prediction task,the multi-mode feature fusion model showed no significant differences in AUC compared to single-mode WSI model(Z=1.739,P=0.082),however its AUC was significantly higher than single-mode T,-FLAIR model(Z=4.830,P=0.000).Conclusions Multi-mode fusion of macro-and micro-imaging features improves prediction accuracy for IDHI genotyping and WHO grading in gliomas,providing a reliable artificial intelligence(Al)decision-support tool for personalized clinical management.

关 键 词：神经胶质瘤 磁共振成像 病理学 基因 肿瘤分级 深度学习 ROC曲线 

分 类 号：R739.41[医药卫生—肿瘤]