《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (9): 2996-3004.DOI: 10.11772/j.issn.1001-9081.2025080954

• 多媒体计算与计算机仿真 • 上一篇    

基于层次化聚合与提示增强的肾小球肾炎分类

张桂滨1,2,3, 米昀1,2,3, 路艳蒙4, 耿舰5,6, 周志涛4, 曹蕾1,2,3()   

  1. 1.南方医科大学 生物医学工程学院,广州 510515
    2.广东省医学图像处理重点实验室(南方医科大学),广州 510515
    3.广东省医学成像与诊断技术工程实验室(南方医科大学),广州 510515
    4.南方医科大学 中心实验室,广州 510515
    5.南方医科大学 基础医学院,广州 510515
    6.南方医科大学 广州华银医学检验中心,广州 510515
  • 收稿日期:2025-08-19 修回日期:2025-10-09 接受日期:2025-10-16 发布日期:2025-11-07 出版日期:2026-09-10
  • 通讯作者: 曹蕾
  • 作者简介:张桂滨(2002—),男,广东汕头人,硕士研究生,主要研究方向:医学图像处理、多模态学习
    米昀(2001—),男,湖南怀化人,硕士研究生,主要研究方向:图像分类、医学人工智能
    路艳蒙(1970—),女,内蒙古呼和浩特人,副教授,博士,主要研究方向:组织胚胎学
    耿舰(1972—),男,广东广州人,副教授,博士,主要研究方向:肾活检病理诊断
    周志涛(1975—),男,辽宁沈阳人,高级实验师,博士,主要研究方向:免疫学
    曹蕾(1974—),女,湖南沅江人,教授,博士,主要研究方向:医学图像处理、深度学习。
  • 基金资助:
    国家自然科学基金资助项目(32071368)

Glomerulonephritis classification based on hierarchical aggregation and prompt enhancement

Guibin ZHANG1,2,3, Yun MI1,2,3, Yanmeng LU4, Jian GENG5,6, Zhitao ZHOU4, Lei CAO1,2,3()   

  1. 1.School of Biomedical Engineering,Southern Medical University,Guangzhou Guangdong 510515,China
    2.Guangdong Provincial Key Laboratory of Medical Image Processing (Southern Medical University),Guangzhou Guangdong 510515,China
    3.Guangdong Provincial Engineering Laboratory for Medical Imaging and Diagnostic Technology (Southern Medical University),Guangzhou Guangdong 510515,China
    4.Central Laboratory,Southern Medical University,Guangzhou Guangdong 510515,China
    5.School of Basic Medical Sciences,Southern Medical University,Guangzhou Guangdong 510515,China
    6.Guangzhou Huayin Medical Laboratory Center,Southern Medical University,Guangzhou Guangdong 510515,China
  • Received:2025-08-19 Revised:2025-10-09 Accepted:2025-10-16 Online:2025-11-07 Published:2026-09-10
  • Contact: Lei CAO
  • About author:ZHANG Guibin, born in 2002, M. S. candidate. His research interests include medical image processing, multimodal learning.
    MI Yun, born in 2001, M. S. candidate. His research interests include image classification, medical artificial intelligence.
    LU Yanmeng, born in 1970, Ph. D., associate professor. Her research interests include histoembryology.
    GENG Jian, born in 1972, Ph. D., associate professor. His research interests include renal biopsy pathological diagnosis.
    ZHOU Zhitao, born in 1975, Ph. D., senior experimentalist. His research interests include immunology.
    CAO Lei, born in 1974, Ph. D., professor. Her research interests include medical image processing, deep learning.
  • Supported by:
    National Natural Science Foundation of China(32071368)

摘要:

肾小球肾炎(GN)亚型的精准判别是临床决策的关键依据,然而,GN的自动化诊断在建模包内非连续透射电子显微镜(TEM)图像间的内在关联与有效利用病理诊断报告中的细粒度专家知识方面存在挑战。因此,提出一种包含层次化Transformer聚合(HTA)模块与报告记忆增强(RME)模块的图文多模态模型CLIP-HaRe(CLIP framework with HTA and RME)。HTA模块采用“子图-示例-包”的分层聚合策略,即通过自注意力机制建模非连续TEM图像间的语义关联,并用最大池化提取关键病变特征;RME模块基于现代Hopfield网络构建病理报告记忆库,利用大语言模型(LLM)生成的类别提示进行查询,并融合报告的细粒度文本知识,增强提示的疾病特异性表征能力。在所构建的包含3种GN亚型的数据集上的实验结果显示,所提模型在疾病分类任务中的准确率、精确率、召回率、F1分数和受试者工作特征曲线下面积(AUC)分别为0.773 7、0.781 6、0.772 3、0.772 7和0.930 8,相较于基线模型线性探测(Linear-Probe),上述指标分别提升了11.01、10.68、10.69、11.04和9.85个百分点。CLIP-HaRe通过融合图文信息实现了GN亚型的自动分类,它的性能相较于基线明显提升,验证了它在辅助临床诊断中的应用潜力。

关键词: 层次化聚合, 提示增强, 图文多模态, 透射电子显微镜图像, 肾小球肾炎

Abstract:

Accurately distinguishing GlomeruloNephritis (GN) subtypes is crucial for clinical decision-making, however, GN automated diagnosis faces challenges in modeling the intrinsic relationships between non-consecutive Transmission Electron Microscopy (TEM) images within a bag and utilizing the fine-grained expert knowledge in pathology reports effectively. Therefore, an image-text multimodal model, CLIP-HaRe (CLIP framework with HTA and RME), was proposed, comprising a Hierarchical Transformer Aggregation (HTA) module and a Report Memory Enhancement (RME) module. In HTA module, a hierarchical aggregation strategy “sub-image-instance-bag” was utilized, in which a self-attention mechanism was employed to model the semantic relationships among non-consecutive TEM images, and max-pooling was used to extract key lesion features. In RME module, the pathology report memory banks were constructed using modern Hopfield network, the class prompts generated by a Large Language Model (LLM) were employed for query, and fine-grained textual knowledge was integrated from the reports, thereby enhancing the disease-specific representation capability of the prompts. Experimental results in the disease classification task on a constructed dataset of three GN subtypes show that the proposed model achieves an accuracy of 0.773 7, a precision of 0.781 6, a recall of 0.772 3, an F1 score of 0.772 7, and an Area Under the receiver operating Characteristic curve (AUC) of 0.930 8. These metrics represent improvements of 11.01, 10.68, 10.69, 11.04, and 9.85 percentage points, respectively, over a baseline method Linear-Probe. CLIP-HaRe enables the automated classification of GN subtypes by fusing image and text information, with its significant performance enhancement over the baseline, validating its potential for application in assisting clinical diagnosis.

Key words: hierarchical aggregation, prompt enhancement, image-text multimodal, Transmission Electron Microscopy (TEM) image, GlomeruloNephritis (GN)

中图分类号: