《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (10): 3371-3380.DOI: 10.11772/j.issn.1001-9081.2024101442

• 前沿与综合应用 • 上一篇    

基于证据深度学习的不确定性感知无监督医学图像配准模型

王一铭1, 李世源1, 廖南清2, 陈庆锋1()   

  1. 1.广西大学 计算机与电子信息学院,南宁 530004
    2.广西大学 医学院,南宁 530004
  • 收稿日期:2024-10-11 修回日期:2024-12-30 接受日期:2024-12-31 发布日期:2025-01-06 出版日期:2025-10-10
  • 通讯作者: 陈庆锋
  • 作者简介:王一铭(2000—),男,广西南宁人,硕士研究生,主要研究方向:医学图像处理、可信机器学习、数据挖掘
    李世源(2000—),男(苗族),湖南沅陵人,硕士研究生,主要研究方向:可信机器学习、图神经网络、图数据挖掘
    廖南清(1990—),男(壮族),广西南宁人,博士,主要研究方向:医学图像处理、生物学信息学、多模态数据融合
    陈庆锋(1972—),男,广西鹿寨人,教授,博士,主要研究方向:医学图像处理、生物信息学、数据挖掘、知识图谱。 Email:qingfeng@gxu.edu.cn
  • 基金资助:
    广西科技基地和人才专项(桂科AD24010011);广西研究生教育创新计划项目(YCSW2024138)

Uncertainty-aware unsupervised medical image registration model based on evidential deep learning

Yiming WANG1, Shiyuan LI1, Nanqing LIAO2, Qingfeng CHEN1()   

  1. 1.School of Computer,Electronics and Information,Guangxi University,Nanning Guangxi 530004,China
    2.Medicine School,Guangxi University,Nanning Guangxi 530004,China
  • Received:2024-10-11 Revised:2024-12-30 Accepted:2024-12-31 Online:2025-01-06 Published:2025-10-10
  • Contact: Qingfeng CHEN
  • About author:WANG Yiming, born in 2000, M. S. candidate. His research interests include medical image processing, trustworthy machine learning, data mining.
    LI Shiyuan, born in 2000, M. S. candidate. His research interests include trustworthy machine learning, graph neural network, graph data mining.
    LIAO Nanqing, born in 1990, Ph. D. His research interests include medical image processing, bioinformatics, multimodal data fusion.
    CHEN Qingfeng, born in 1972, Ph. D., professor. His research interests include medical image processing, bioinformatics, data mining, knowledge graph.
  • Supported by:
    Innovation Program of Guangxi Graduate Education(YCSW2024138);Guangxi Science and Technology Base and Talent Specific Project(GuiKe AD24010011)

摘要:

医学图像配准中的不确定性量化对医师在实际临床应用中评估风险至关重要。目前,基于深度无监督学习的医学图像配准模型虽然已经具备不错的效果,但仍缺乏在配准时估计外观不确定性的方法,这将影响配准的精度和可信度。此外,在实时性应用场景中,医学图像配准模型不但需要具备较高的配准精度,还需要快速进行推理。针对上述问题,提出一种基于证据深度学习(EDL)的不确定性感知无监督医学图像配准模型EvidentialMorph,将EDL应用于无监督医学图像配准。EDL是一种不确定性量化方法,无需额外的计算开销。首先,通过U-net架构的配准主干网络模块学习得到形变向量场(DVF);其次,通过一种改进的空间变换器网络(STN)模块——证据STN模块学习配准图像的正态逆伽马(NIG)分布,从而直接计算出配准图像及其外观不确定性。在海马体、LPBA40和IBSR18核磁共振成像(MRI)数据集上的实验结果表明,与CLMorph模型相比,在配准精度上,EvidentialMorph的Dice相似性系数(DSC)最高提升了3.31%,归一化交叉相关(NCC)系数最高提升了2.75%;在推理时耗上,EvidentialMorph减少了85 ms。以上结果验证了EvidentialMorph能够快速得到有效的不确定性量化结果,并提升配准精度,为实时医学图像配准场景提供了不确定性量化的可能,可以改善配准效果。

关键词: 医学图像配准, 无监督学习, 不确定性估计, 证据深度学习, 可信机器学习

Abstract:

Uncertainty quantification in medical image registration is crucial for doctors to evaluate risk in real-world clinical applications. Recently, deep unsupervised learning-based medical image registration models have shown certain effects, but there is a lack of methods to estimate appearance uncertainty during registration, which will affect registration accuracy and trustworthiness. In addition, in real-time application scenarios, medical image registration models need to be highly accurate and fast in inference at the same time, which is difficult to be achieved by the existing models. To address these issues, an uncertainty-aware unsupervised medical image registration model based on Evidential Deep Learning (EDL) — EvidentialMorph was proposed to apply EDL, an uncertainty quantification approach without additional computational cost, to unsupervised medical image registration. Firstly, the Deformation Vector Field (DVF) was learnt and obtained through a registration backbone network module with a U-net architecture. Then, the Normal-Inverse Gamma (NIG) distribution of the registered image was learnt and obtained through an improved Spatial Transformer Network (STN) module — evidential STN module, thereby calculating the registered image and its appearance uncertainty directly. Experiments were carried out on Hippocampus, LPBA40, and IBSR18 Magnetic Resonance Imaging (MRI) datasets. The results show that in registration accuracy, EvidentialMorph improves the Dice Similarity Coefficient (DSC) and Normalized Cross-Correlation (NCC) coefficient by 3.31% and 2.75% at most, respectively, over CLMorph model; and in inference time, EvidentialMorph reduces 85 ms. The above results verify that EvidentialMorph can obtain effective uncertainty quantification quickly and improves registration accuracy, offering potential for real-time medical image registration scenarios and improving registration effects.

Key words: medical image registration, unsupervised learning, uncertainty estimation, Evidential Deep Learning (EDL), trustworthy machine learning

中图分类号: