基于图神经网络的胆囊癌CT影像辅助诊断模型

doi:10.3969/j.issn.1674-8115.2025.09.014

上海交通大学学报（医学版） ›› 2025, Vol. 45 ›› Issue (9): 1221-1231.doi: 10.3969/j.issn.1674-8115.2025.09.014

• 论著 · 技术与方法 • 上一篇下一篇

基于图神经网络的胆囊癌CT影像辅助诊断模型

尹梓名¹, 王荣钦¹, 杨自逸², 刘颖斌³, 陈涛³, 束翌俊², 龚伟²()

^1.上海理工大学健康科学与工程学院，上海 200093
^2.上海交通大学医学院附属新华医院普外科，上海 200092
^3.上海交通大学医学院附属仁济医院胆胰外科，上海 200127

收稿日期:2025-06-09 接受日期:2025-08-25 出版日期:2025-09-28 发布日期:2025-09-30
通讯作者: 龚　伟，主任医师，博士；电子信箱：gongwei@xinhuamed.com.cn。
基金资助:
国家自然科学基金面上项目(82173081);上海市卫生健康委员会卫生健康青年人才(2022YQ002);上海交通大学医学院“双百人”项目(20151001)

Graph neural network-based auxiliary diagnostic model for gallbladder cancer on CT imaging

YIN Ziming¹, WANG Rongqin¹, YANG Ziyi², LIU Yingbin³, CHEN Tao³, SHU Yijun², GONG Wei²()

^1.School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
^2.Department of General Surgery, Xinhua Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai 200092, China
^3.Department of Hepatobiliary and Pancreatic Surgery, Renji Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai 200127, China

Received:2025-06-09 Accepted:2025-08-25 Online:2025-09-28 Published:2025-09-30
Contact: GONG Wei, E-mail: gongwei@xinhuamed.com.cn.
Supported by:
National Natural Science Foundation of China(82173081);Young Talents in Health of Shanghai Municipal Health Commission(2022YQ002);“Two-hundred Talents” Program of Shanghai Jiao Tong University School of Medicine(20151001)

摘要/Abstract

摘要：

目的·建立一种基于图神经网络（graph neural network，GNN）的胆囊癌CT影像辅助诊断模型，并验证其准确性与可行性。方法·收集2010年1月至2023年11月上海交通大学医学院附属新华医院和附属仁济医院887例正常胆囊、胆囊良性疾病、胆囊癌患者的1 774张腹部增强CT动脉期影像，按照4∶1随机划分为训练集和测试集，建立GNN和卷积神经网络（convolutional neural network，CNN）混合架构的模型VJK-GIN。该模型通过像素级的图构建方式，将每个像素作为图中的一个节点，空间邻接为边，从中提取图像局部纹理特征。在模型结构设计上，VJK-GIN模型采用3层图同构网络模块，并引入虚拟节点模块和跳跃知识机制，最后采用全局池化操作将节点特征压缩为图表示，并通过多层感知机分类器输出诊断结果。通过五折交叉验证法，比较VJK-GIN模型与其他GNN（GCN、GraphSAGE、GAT和GIN）、CNN（ViT、EfficientNetV2和ConvNeXt）模型的准确率、精确度、召回率、F1分数以及受试者操作特征曲线（receiver operator characteristic curve，ROC曲线）下面积（area under curve，AUC）。结果·五折交叉验证法结果显示，VJK-GIN模型的F1分数为0.799（95%CI 0.775~0.823），召回率为0.795（95%CI 0.773~0.817），精确度为0.799（95%CI 0.775~0.823），AUC为0.812（95%CI 0.792~0.832），准确率为0.773（95%CI 0.748~0.798），所有评估指标均优于其他模型。结论·VJK-GIN模型对于正常胆囊、胆囊良性疾病、胆囊癌增强CT图像识别的稳定性和准确性均较高。

关键词: 胆囊癌, 图神经网络, 卷积神经网络, 医学影像分析, 深度学习

Abstract:

Objective ·To develop a graph neural network (GNN)-based auxiliary diagnostic model for gallbladder cancer on CT images, and validate its accuracy and feasibility. Methods ·From January 2010 to November 2023, 1 774 contrast-enhanced CT arterial-phase images were acquired from 887 patients with normal gallbladder, benign gallbladder disease, or gallbladder cancer at Xinhua Hospital and Renji Hospital, Shanghai Jiao Tong University School of Medicine. These images were randomly divided into training and testing sets at a 4∶1 ratio to develop a hybrid GNN-convolutional neural network (CNN) model, named VJK-GIN. The model constructed a pixel-level graph in which each pixel served as a node, and spatial adjacency defined the edges, enabling extraction of local texture features. In the model architecture design, VJK-GIN integrated a three-layer graph isomorphism network, augmented with virtual nodes and jump-knowledge connections; global pooling compressed node features into a graph-level representation, which was classified by a multi-layer perceptron head. Five-fold cross-validation was used to compare VJK-GIN with GNN baselines (GCN, GraphSAGE, GAT, and GIN) and CNN baselines (ViT, EfficientNetV2, and ConvNeXt) in terms of accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Results ·The results of five-fold cross-validation showed that VJK-GIN achieved an F1-score of 0.799 (95%CI 0.775‒0.823), recall of 0.795 (95%CI 0.773‒0.817), precision of 0.799 (95%CI 0.775‒0.823), AUC of 0.812 (95%CI 0.792‒0.832), and accuracy of 0.773 (95%CI 0.748‒0.798), surpassing all competing models across every metric. Conclusion ·The VJK-GIN model exhibits high stability and accuracy in identifying contrast-enhanced CT images of normal, benign, and malignant gallbladder conditions.

Key words: gallbladder cancer, graph neural network (GNN), convolutional neural network (CNN), medical image analysis, deep learning

中图分类号:

R735.8

尹梓名, 王荣钦, 杨自逸, 刘颖斌, 陈涛, 束翌俊, 龚伟. 基于图神经网络的胆囊癌CT影像辅助诊断模型[J]. 上海交通大学学报（医学版）, 2025, 45(9): 1221-1231.

YIN Ziming, WANG Rongqin, YANG Ziyi, LIU Yingbin, CHEN Tao, SHU Yijun, GONG Wei. Graph neural network-based auxiliary diagnostic model for gallbladder cancer on CT imaging[J]. Journal of Shanghai Jiao Tong University (Medical Science), 2025, 45(9): 1221-1231.

图/表 7

图1 数据纳入流程

Fig 1 Process of data inclusion

图2 全自动胆囊区域识别裁剪算法流程图

Fig 2 Flow chart of the fully automatic gallbladder region recognition and clipping algorithm

图3 胆囊癌诊断模型结构示意图Note： ROI—region of interest; k-NN—k-nearest neighbors; MLP—multi-layer perceptron.

Fig 3 Overall schematic diagram of the model for gallbladder cancer diagnosis

表1 训练集与测试集基线特征比较（五折交叉验证中其中一折）

Tab 1 Comparison of baseline characteristics between the training and test sets （one fold of 5-fold cross-validation）

Item	Total	Training set	Testing set	P value
Number of patients/n	887	709	178	‒
Number of CT slices/n	1 774	1419	355	‒
Male/n (%)	447 (50.4)	360 (50.8)	87 (49.2)	0.712
Age/year	64.6±10.4	65.0±10.6	63.0±9.6	0.331
BMI/(kg·m^-2)	24.3±3.1	24.4±3.5	23.9±0.18	0.383
Disease type/n (%)				>0.999
Cancer	266 (30)	213 (30)	53 (30)
Benign	266 (30)	213 (30)	53 (30)
Normal	355 (40)	283 (40)	72 (40)

表2 各模型主要指标对比

Tab 2 Comparison of main indicators of each model

Recall

(95%CI)

0.795

(0.773‒0.817)

0.668

(0.641‒0.695)

0.713

(0.689‒0.737)

0.749

(0.729‒0.769)

0.760

(0.744‒0.776)

0.777

(0.760‒0.795)

0.752

(0.729‒0.776)

0.620

(0.594‒0.645)

Precision

(95%CI)

0.799

(0.775‒0.823)

0.667

(0.638‒0.696)

0.715

(0.695‒0.735)

0.742

(0.721‒0.764)

0.758

(0.743‒0.776)

0.778

(0.759‒0.798)

0.754

(0.730‒0.778)

0.621

(0.593‒0.648)

F1-score

(95%CI)

0.799

(0.775‒0.823)

0.667

(0.638‒0.696)

0.714

(0.694‒0.734)

0.742

(0.721‒0.764)

0.758

(0.743‒0.776)

0.778

(0.758‒0.797)

0.753

(0.729‒0.776)

0.619

(0.592‒0.646)

Accuracy

(95%CI)

0.773

(0.748‒0.798)

0.668

(0.644‒0.692)

0.713

(0.695‒0.731)

0.728

(0.708‒0.748)

0.740

(0.718‒0.762)

0.777

(0.758‒0.797)

0.752

(0.731‒0.774)

0.620

(0.592‒0.647)

AUC

(95%CI)

0.812

(0.792‒0.832)

0.697

(0.672‒0.722)

0.713

(0.691‒0.735)

0.762

(0.742‒0.782)

0.775

(0.755‒0.795)

0.757

(0.735‒0.778)

0.751

(0.727‒0.774)

0.703

(0.677‒0.728)

图4 各模型在测试集中的混淆矩阵热力图Note： Color bars represent the intensity of the models' prediction, with darker shades corresponding to more accurate predictions. Numbers represent the counts of true positives, false positives, and false negatives for each class. The diagonal values are the true positives (correct predictions), while off-diagonal values are misclassifications (false positives and false negatives).

Fig 4 Confusion matrix heatmaps of each model in the test set

图5 各模型的综合性能可视化分析Note： A. ROC curves. B. Accuracy curves of the training set. C. Loss curves of the training set. D/E. Bar chart (D) and radar chart (E) of 5 indexes by five-fold cross-validation. TPR—true positive rate; FPR—false positive rate.

Fig 5 Visualization of the comprehensive performance of each model

参考文献 32

[1]	刘颖斌, 陈炜. 重视胆囊癌的临床和基础研究[J]. 中华消化外科杂志, 2025, 24(1): 64-71.
	LIU Y B, CHEN W. Attach importance to the clinical and basic research on gallbladder carcinoma[J]. Chinese Journal of Digestive Surgery, 2025, 24(1): 64-71.
[2]	刘珊山, 李初谊, 郑英, 等. 1990—2019年中国胆囊癌疾病负担研究及未来趋势分析[J]. 中国全科医学, 2024, 27(14): 1742-1749.
	LIU S S, LI C Y, ZHENG Y, et al. Disease burden of gallbladder cancer in China from 1990 to 2019 and the analysis of its future trends[J]. Chinese General Practice, 2024, 27(14): 1742-1749.
[3]	吴玉梅, 郝金钢. 影像组学在胆囊癌诊疗中的研究进展[J]. 国际医学放射学杂志, 2024, 47(5): 594-598.
	WU Y M, HAO J G. Research progress of radiomics in the diagnosis and treatment of gallbladder cancer[J]. International Journal of Medical Radiology, 2024, 47(5): 594-598.
[4]	罗明伟, 张译中, 周田, 等. 意外胆囊癌规范化诊疗策略的研究进展[J]. 腹部外科, 2022, 35(6): 457-461.
	LUO M W, ZHANG Y Z, ZHOU T, et al. Research advances on standardized diagnosis and treatment strategies for unexpected gallbladder cancer[J]. Journal of Abdominal Surgery, 2022, 35(6): 457-461.
[5]	马振威, 朱博, 刘赋斌, 等. 血小板和淋巴细胞比值联合CA19-9在胆囊癌术后患者预后评估中的价值[J]. 中华肝脏外科手术学电子杂志, 2024, 13(2): 163-168.
	MA Z W, ZHU B, LIU F B, et al. Evaluation value of platelet-to-lymphocyte ratio combined with CA19-9 for postoperative prognosis of patients with gallbladder cancer[J]. Chinese Journal of Hepatic Surgery (Electronic Edition), 2024, 13(2): 163-168.
[6]	杨志新, 赵丽珠, 邓玥, 等. 基于卷积神经网络在卵巢肿瘤良恶性鉴别诊断中的研究[J]. 昆明医科大学学报, 2023, 44(10): 134-139.
	YANG Z X, ZHAO L Z, DENG Y, et al. Differential diagnosis of benign and malignant ovarian tumors based on convolutional neural network[J]. Journal of Kunming Medical University, 2023, 44(10): 134-139.
[7]	沈天乐, 杜向慧. 人工智能在恶性肿瘤放疗领域中的应用与前景[J]. 浙江医学, 2018, 40(8): 783-785, 795.
	SHEN T L, DU X H. Application and prospect of artificial intelligence in radiotherapy for malignant tumors[J]. Zhejiang Medicine, 2018, 40(8): 783-785, 795.
[8]	蒋西然, 蒋韬, 孙嘉瑶, 等. 深度学习人工智能技术在医学影像辅助分析中的应用[J]. 中国医疗设备, 2021, 36(6): 164-171.
	JIANG X R, JIANG T, SUN J Y, et al. Deep learning in computer aided analyses of medical images[J]. China Medical Devices, 2021, 36(6): 164-171.
[9]	张礼, 王嘉瑞. 基于图同构网络的自闭症功能磁共振影像诊断算法[J]. 南京大学学报(自然科学), 2021, 57(5): 801-809.
	ZHANG L, WANG J R. Diagnosing autism spectrum disorder from functional MRI using graph isomorphic network[J]. Journal of Nanjing University (Natural Sciences), 2021, 57(5): 801-809.
[10]	覃智威, 刘钊, 陆允敏, 等. 基于广义卷积神经网络的阿尔茨海默病多模态磁共振图像分类方法研究[J]. 生物医学工程学杂志, 2023, 40(2): 217-225.
	QIN Z W, LIU Z, LU Y M, et al. Research on classification method of multimodal magnetic resonance images of Alzheimer's disease based on generalized convolutional neural networks[J]. Journal of Biomedical Engineering, 2023, 40(2): 217-225.
[11]	JEONG Y, KIM J H, CHAE H D, et al. Deep learning-based decision support system for the diagnosis of neoplastic gallbladder polyps on ultrasonography: preliminary results[J]. Sci Rep, 2020, 10(1): 7700.
[12]	ZHAO H Y, MIAO C, ZHU Y D, et al. An end-to-end interpretable machine-learning-based framework for early-stage diagnosis of gallbladder cancer using multi-modality medical data[J]. BMC Cancer, 2025, 25(1): 1178.
[13]	YANG M, ZHAO Y H, LI C, et al. Multimodal integration of liquid biopsy and radiology for the noninvasive diagnosis of gallbladder cancer and benign disorders[J]. Cancer Cell, 2025, 43(3): 398-412.e4.
[14]	YIN Y C, YAKAR D, SLANGEN J J G, et al. The value of deep learning in gallbladder lesion characterization[J]. Diagnostics (Basel), 2023, 13(4): 704.
[15]	GUPTA P, BASU S, RANA P, et al. Deep-learning enabled ultrasound based detection of gallbladder cancer in northern India: a prospective diagnostic study[J]. Lancet Reg Health Southeast Asia, 2024, 24: 100279.
[16]	OBAID A M, TURKI A, BELLAAJ H, et al. Detection of gallbladder disease types using deep learning: an informative medical method[J]. Diagnostics (Basel), 2023, 13(10): 1744.
[17]	WANG H X, HUANG G, ZHAO Z, et al. CCF-GNN: a unified model aggregating appearance, microenvironment, and topology for pathology image classification[J]. IEEE Trans Med Imaging, 2023, 42(11): 3179-3193.
[18]	ZHANG L, ZHAO Y, CHE T T, et al. Graph neural networks for image-guided disease diagnosis: a review[J]. iRADIOLOGY, 2023, 1(2): 151-166.
[19]	ADDOUNE S E, OUAHBI M I. A hybrid UNet-GNN architecture for enhanced medical image segmentation[D]. Ouargla, Algeria: Kasdi Merbah University, 2024.
[20]	DUAN J W, XIONG J Q, LI Y H, et al. Deep learning based multimodal biomedical data fusion: an overview and comparative review[J]. Inf Fusion, 2024, 112: 102536.
[21]	LING R H, CEN X X, WU S Y, et al. Explainable graph neural network based on metabolic brain imaging for differential diagnosis of Parkinsonism[J]. Front Aging Neurosci, 2025, 17: 1580910.
[22]	ZHENG K Z, YU S J, CHEN L J, et al. BPI-GNN: interpretable brain network-based psychiatric diagnosis and subtyping[J]. Neuroimage, 2024, 292: 120594.
[23]	XU K, HU W, LESKOVEC J, et al. How powerful are graph neural networks?[DB/OL]. (2019-02-22)[2025-08-03]. https://arxiv.org/abs/1810.00826.
[24]	WANG H, GAO K, ZHANG X, et al. Pixel- and patch-wise context-aware learning with CNN and GCN collaboration for hyperspectral image classification[C]//IGARSS 2023‒2023 IEEE International Geoscience and Remote Sensing Symposium, Pasadena, CA, USA. New York: IEEE, 2023: 7555-7558.
[25]	WANG Y F, YU X D, DONG H B, et al. Graph convolutional network with local and global feature fusion for hyperspectral image classification[J]. IEEE Trans Geosci Remote Sens, 2024, 62: 5536915.
[26]	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA. New York: IEEE, 2016: 770-778.
[27]	DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16×16 words: transformers for image recognition at scale[DB/OL]. (2021-06-03)[2025-08-03]. https://arxiv.org/abs/2010.11929.
[28]	TAN M, LE Q V. EfficientNetV2: smaller models and faster training[DB/OL]. (2021-06-23)[2025-08-03]. https://arxiv.org/abs/2104.00298.
[29]	FENG J W, TAN H L, LI W W, et al. Conv2NeXt: reconsidering ConvNeXt network design for image recognition[C]//2022 International Conference on Computers and Artificial Intelligence Technologies (CAIT), Quzhou, China. New York: IEEE, 2022: 53-60.
[30]	KIPF T N, WELLING M. Semi-supervised classification with graph convolutional networks[DB/OL]. (2017-02-22)[2025-08-03]. https://arxiv.org/abs/1609.02907.
[31]	HAMILTON W L, YING R, LESKOVEC J. Inductive representation learning on large graphs[DB/OL]. (2018-09-10)[2025-08-03]. https://arxiv.org/abs/1706.02216.
[32]	VELIČKOVIĆ P, CUCURULL G, CASANOVA A, et al. Graph attention networks[DB/OL]. (2018-02-04)[2025-08-03]. https://arxiv.org/abs/1710.10903.

基于图神经网络的胆囊癌CT影像辅助诊断模型

Graph neural network-based auxiliary diagnostic model for gallbladder cancer on CT imaging

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 7

参考文献 32

相关文章 11

编辑推荐

Metrics

[1]	陈健, 王珍妮, 夏开建, 王甘红, 刘罗杰, 徐晓丹. 基于深度学习的结直肠息肉内镜图像分割和分类方法比较[J]. 上海交通大学学报（医学版）, 2024, 44(6): 762-772.
[2]	刘洋, 吴梦怡, 胡尧, 亓坤, 王渝彬, 赵悦, 宋锦璘. 基于Transformer和扩散模型的头颅侧位片颈椎分割方法在正畸临床中的初步应用[J]. 上海交通大学学报（医学版）, 2024, 44(12): 1579-1586.
[3]	杨婧潇, 贾子尧, 吴文广, 吴向嵩, 张飞, 李怀峰, 朱逸荻, 李茂岚. BRCA1 R1325K突变对胆囊癌细胞增殖及凋亡的影响[J]. 上海交通大学学报（医学版）, 2023, 43(9): 1071-1079.
[4]	赵珂珂, 蒋蓓蓓, 张璐, 王凌云, 张亚平, 解学乾. 超低剂量平扫CT深度学习图像重建评价肺部病灶的可行性[J]. 上海交通大学学报（医学版）, 2022, 42(8): 1062-1069.
[5]	林嘉希, 汪盛嘉, 赵鑫, 高欣, 殷民月, 朱锦舟. 基于深度卷积神经网络的Barrett食管内镜图片分类模型的建立[J]. 上海交通大学学报（医学版）, 2022, 42(5): 653-659.
[6]	陈立奇, 薛卓维, 吴氢凯. 基于磁共振成像的女性盆底器官三维数字模型重建的研究进展[J]. 上海交通大学学报（医学版）, 2022, 42(3): 381-386.
[7]	王晓峰, 周璐, 姚乐宇, 何凡, 彭海霞, 杨大明, 黄晓霖. 深度卷积神经网络模型辅助下结直肠息肉检测系统对初级医师结直肠小息肉检出率的影响[J]. 上海交通大学学报(医学版), 2022, 42(2): 205-210.
[8]	李小敏, 曲扬, 张少霆, 赵亮, 刘畅, 谢帅宁, 戴尅戎, 艾松涛. 人工智能技术在骨肌系统影像学方面的应用[J]. 上海交通大学学报(医学版), 2021, 41(2): 262-266.
[9]	陆言巧，沈兰，何奔. 人工智能在心血管疾病的辅助诊疗中的应用[J]. 上海交通大学学报（医学版）, 2020, 40(2): 259-.
[10]	蒋蓓蓓，张亚平，张琳，刘桂雪，解学乾. 深度卷积神经网络对≤ 3 cm的亚实性肺腺癌 CT图像病理学分型预测的可视化研究[J]. 上海交通大学学报（医学版）, 2019, 39(9): 1045-.
[11]	秦一雨，周学平，林培艺，等. 吉西他滨联合索拉非尼对胆囊癌细胞生长侵袭的抑制作用[J]. 上海交通大学学报（医学版）, 2014, 34(3): 318-.