基于多约束潜在表征学习的双向特征映射分类模型：肺炎鉴别诊断

doi:10.12122/j.issn.1673-4254.2026.06.23

摘要/Abstract

摘要：

目的构建一种基于多约束潜在表征学习的双向特征映射肺炎鉴别分类模型，并验证其判别性能、可解释性及临床可行性。方法收集了2913名患者（1457名阳性和1456名阴性）的胸部X线（CXR）图像，并使用开源影像组学工具PyRadiomics从掩码中提取影像组学特征。使用本研究提出的基于多约束潜在表征学习的双向特征映射分类模型，将CXR的影像组学特征映射到潜在共享空间并建立分类模型。采用五折交叉验证方法和阳性预测值（PPV）、阴性预测值（NPV）、特异性（SPE）、灵敏度（SEN）、准确率（ACC）、ROC曲线下面积（AUC）评价该分类模型的鉴别性能。将本研究所提出的模型与其他特征分类模型对于肺炎的鉴别能力进行定量比较。决策曲线分析（DCA）和消融实验评估各约束模块的贡献。采用 SHAP 方法对影像组学特征的重要性进行解释，以提升模型的可解释性。本研究提出特征映射方法得到的低维潜在特征进行样本散点可视化实验，验证本研究所提出的特征映射分类模型的可行性和有效性。结果五折交叉验证结果显示本研究所提出的基于多约束表征学习的双向特征映射分类模型在鉴别肺炎中的PPV、NPV、SPE、SEN、ACC、AUC分别为0.796、0.830、0.784、0.838、0.811、0.893，决策曲线显示在临床可接受阈值内具有更高净受益，消融实验验证多约束模块的关键作用，SHAP 分析表明模型关注的特征具有医学合理性，且特征映射方法在可视化实验中具有优秀的表现。结论基于多约束潜在表征学习的双向特征映射分类模型在鉴别肺炎中的应用具有较强的鉴别能力和较高的应用价值。与其他分类模型相比，本研究提出的分类模型在肺炎的鉴别分类任务中具有较大的优势。

关键词: 影像组学, 特征映射, 表征学习, 判别分析, 肺炎

Abstract:

Objective To develop a bidirectional feature-mapping classification model for differential diagnosis of pneumonia. Methods We collected chest X-ray (CXR) images from 1457 patients with pneumonia and 1456 healthy individuals. Radiomic features extracted from the segmentation masks using PyRadiomics were mapped into a latent shared space to construct the classification model using the bidirectional feature mapping classification model based on multi-constrained latent representation learning. The performance of the constructed model for differential diagnosis of pneumonia was evaluated using 5-fold cross-validation and compared with other feature-based classification models. Decision curve analysis was used for evaluating clinical utility of the model, and ablation experiments were performed to assess the contribution of each constraint module. The importance of the radiomics features was interpreted using the SHAP method, and a two-dimensional visualization experiment of the low-dimensional latent features obtained through the proposed mapping method was conducted to verify the feasibility and effectiveness of the model. Results The 5-fold cross-validation results showed that the proposed classification model had a positive predictive value of 0.796, a negative predictive value of 0.830, a specificity of 0.784, a sensitivity of 0.830, an accuracy of 0.811, and an area under the ROC curve of 0.893 for differential diagnosis of pneumonia. Decision curve analysis demonstrated a high net clinical benefit of the model within acceptable threshold probabilities. Ablation studies confirmed the essential role of the multi-constraint module, and the SHAP analysis revealed that the model focused primarily on clinically meaningful and medically interpretable features. The feature mapping method exhibited excellent performance in visual experiments to confirm the effectiveness of the proposed model. Conclusion The proposed bidirectional feature mapping classification model demonstrates strong discriminative capability and high potential for differential diagnosis of pneumonia and shows obvious advantages over other classification models in pneumonia classification tasks.

Key words: radiomics, feature mapping, representation learning, discriminant analysis, pneumonia

曾敏, 卓俐, 谭顺谦, 甄鑫. 基于多约束潜在表征学习的双向特征映射分类模型：肺炎鉴别诊断[J]. 南方医科大学学报, 2026, 46(6): 1434-1443.

Min ZENG, Li ZHUO, Shunqian TAN, Xin ZHEN. A bidirectional feature mapping classification model based on multi-constrained latent representation learning for differential diagnosis of pneumonia[J]. Journal of Southern Medical University, 2026, 46(6): 1434-1443.

图/表 11

图1 基于多约束潜在表征学习的双向特征映射分类模型流程图

Fig.1 Framework of the feature mapping classification model based on multi-constrained latent representation learning.

表1 提取到的影像组学特征

Tab.1 Extracted radiomic features

Types of radiomics features	Radiomics Features
First order features (n=19)	Energy, total energy, entropy minimum, 10th percentile, 90th percentile, maximum, mean, median, interquartile range, range, Mean Absolute Deviation (MAD), Robust Mean Absolute Deviation (rMAD), Root Mean Squared (RMS), Standard deviation, skewness, kurtosis, variance, uniformity
Shape features (n=10)	Elongation, major axis length, maximum diameter, mesh surface, minor axis length, perimeter, perimeter surface ratio, pixel surface, sphericity, spherical disproportion
Texture features (n=74)	¹GLCM (n=24), ²GLRLM (n=16), ³GLSZM (n=16), ⁴NGTDM (n=5), ⁵GLDM (n=14)

表2 特征映射算法伪代码

Tab.2 Pseudocode of the proposed feature mapping algorithm

Algorithm 1 Pseudocode of the Proposed Method

Training stage

Input: data feature matrix $X ∈ R m × n$ ,dimension of potential space k, weight parameters for each constraint $α 1, α 2, α 3, α 4, α 5, α 6$ and $ρ$ .

Output: projection Matrix for Data Representation Learning $Q ∈ R k × m$

Begin

Initialize $C, P, Q, V$ as random matrices;

For t to number of iterations do

Update $C$ by using Eq.(13)

Update $P$ by using Eq.(16)

Update $Q$ by using Eq.(19)

Update $V$ by using Eq.(21)

End

Testing stage

Input: New test dataset $X ̃ ∈ R m × n$

Output: The projection of the new dataset onto the potential space obtain $V n e w = Q X ̃$

表2 特征映射算法伪代码

Tab.2 Pseudocode of the proposed feature mapping algorithm

Algorithm 1 Pseudocode of the Proposed Method

Training stage

Input: data feature matrix $X ∈ R m × n$ ,dimension of potential space k, weight parameters for each constraint $α 1, α 2, α 3, α 4, α 5, α 6$ and $ρ$ .

Output: projection Matrix for Data Representation Learning $Q ∈ R k × m$

Begin

Initialize $C, P, Q, V$ as random matrices;

For t to number of iterations do

Update $C$ by using Eq.(13)

Update $P$ by using Eq.(16)

Update $Q$ by using Eq.(19)

Update $V$ by using Eq.(21)

End

Testing stage

Input: New test dataset $X ̃ ∈ R m × n$

Output: The projection of the new dataset onto the potential space obtain $V n e w = Q X ̃$

表3 采用不同分类器的情况下模型的鉴别性能比较

Tab.3 Performance comparison of the model using different classifiers

Classifier	AUC	ACC	SEN	SPE	PPV	NPV
LDA	0.881	0.800	0.780	0.820	0.814	0.789
KNN	0.854	0.779	0.729	0.828	0.811	0.754
DecisionTree	0.734	0.734	0.733	0.736	0.736	0.734
SVM	0.900	0.813	0.784	0.843	0.834	0.797
RandomForest	0.893	0.805	0.781	0.828	0.821	0.792
XGBoost	0.891	0.804	0.787	0.821	0.816	0.795
MLP	0.898	0.820	0.806	0.834	0.830	0.811
ExtraTree	0.893	0.811	0.838	0.784	0.796	0.830

图2 极端随机树模型在肺炎分类任务中的决策曲线分析结果

Fig.2 Decision curve analysis of the Extra Trees model for pneumonia classification.

表4 不同模块消融设置下模型分类性能比较

Tab.4 Comparison of classification performance of the model using different module ablation settings

Model	AUC	ACC	SEN	SPE	PPV	NPV
Basic Frame	0.821	0.743	0.810	0.675	0.714	0.781
¹w/o Scatter	0.844	0.758	0.787	0.729	0.744	0.775
w/o Soft	0.864	0.780	0.832	0.728	0.755	0.812
w/o Attention	0.855	0.778	0.828	0.728	0.753	0.810
Proposed	0.893	0.811	0.838	0.784	0.796	0.830

表5 基于多约束潜在表征学习的双向特征映射分类模型与其他模型在鉴别肺炎患者中的鉴别性能比较

Tab.5 Performance of the proposed bidirectional feature mapping classification model and other models for pneumonia discrimination

Strategy	AUC	ACC	SEN	SPE	PPV	NPV
Baseline	0.779	0.694	0.773	0.616	0.669	0.732
Laplacian score	0.738	0.681	0.728	0.633	0.666	0.701
UDFS	0.851	0.772	0.802	0.742	0.759	0.790
CIFE	0.839	0.758	0.807	0.709	0.735	0.787
MIFS	0.837	0.758	0.807	0.709	0.736	0.787
mRMR	0.838	0.760	0.810	0.710	0.737	0.790
AutoEncoder	0.876	0.791	0.816	0.765	0.778	0.807
Isomap	0.842	0.762	0.797	0.727	0.746	0.782
Proposed method	0.893	0.811	0.838	0.784	0.796	0.830

表6 基于多约束潜在表征学习的双向特征映射分类模型与其他模型在鉴别肺炎患者中的鉴别性能比较的显著性

Tab.6 Statistics for performance comparison between the proposed bidirectional feature mapping classification model and other models for pneumonia discrimination

Strategy	¹AUC-p	ACC-p	SEN-p	SPE-p	PPV-p	NPV-p
Baseline	1.09E-05	2.58E-05	2.89E-04	2.89E-04	2.06E-05	1.61E-04
Laplacian score	3.38E-05	2.56E-04	2.76E-03	1.34E-04	1.60E-04	3.13E-02
UDFS	6.86E-03	2.41E-03	1.96E-02	1.28E-02	3.37E-03	8.33E-03
CIFE	2.06E-04	3.92E-04	1.74E-02	1.92E-03	1.28E-03	2.08E-03
MIFS	3.36E-04	1.13E-03	1.07E-02	1.97E-03	1.66E-03	3.04E-03
mRMR	1.80E-05	9.52E-04	3.19E-02	1.05E-03	9.09E-04	4.16E-03
AutoEncoder	3.20E-03	1.75E-02	1.68E-02	6.05E-02	2.72E-02	1.60E-02
Isomap	9.81E-05	1.97E-03	5.42E-03	5.19E-03	2.43E-03	2.86E-03
Proposed method	-		-		-	-

图3 原始特征(A)和潜在特征(B)样本散点可视化实验

Fig.3 Scatter plot visualization experiment of original features (A) and latent features (B) samples.

图4 模型的影像组学特征 SHAP 重要性分布图

Fig.4 SHAP summary plot of radiomic feature importance.

图5 所提方法在数据集上的收敛曲线

Fig.5 Convergence curve of the proposed method on a dataset.

参考文献 39

[1]	File TM Jr, Ramirez JA. Community-acquired pneumonia[J]. N Engl J Med, 2023, 389(7): 632-41. doi：10.1056/NEJMcp2303286
[2]	de Benedictis FM, Kerem E, Chang AB, et al. Complicated pneumonia in children[J]. Lancet, 2020, 396(10253): 786-98. doi：10.1016/s0140-6736(20)31550-6
[3]	Theilacker C, Sprenger R, Leverkus F, et al. Population-based incidence and mortality of community-acquired pneumonia in Germany[J]. PLoS One, 2021, 16(6): e0253118. doi：10.1371/journal.pone.0253118
[4]	Rabbah J, Ridouani M, Hassouni L. Improving pneumonia diagnosis with high-accuracy CNN-Based chest X-ray image classification and integrated gradient[J]. Biomed Signal Process Control, 2025, 101: 107239. doi：10.1016/j.bspc.2024.107239
[5]	Sharma S, Guleria K. A systematic literature review on deep learning approaches for pneumonia detection using chest X-ray images[J]. Multimed Tools Appl, 2024, 83(8): 24101-51. doi：10.1007/s11042-023-16419-1
[6]	Koo HJ, Lim S, Choe J, et al. Radiographic and CT features of viral pneumonia[J]. Radiographics, 2018, 38(3): 719-39. doi：10.1148/rg.2018170048
[7]	Fields BKK, Demirjian NL, Dadgar H, et al. Imaging of COVID-19: CT, mri, and pet[J]. Semin Nucl Med, 2021, 51(4): 312-20. doi：10.1053/j.semnuclmed.2020.11.003
[8]	Syrjala H, Broas M, Ohtonen P, et al. Chest magnetic resonance imaging for pneumonia diagnosis in outpatients with lower respiratory tract infection[J]. Eur Respir J, 2017, 49(1): 1601303. doi：10.1183/13993003.01303-2016
[9]	Li YY, Zhang ZY, Dai C, et al. Accuracy of deep learning for automated detection of pneumonia using chest X-Ray images: a systematic review and meta-analysis[J]. Comput Biol Med, 2020, 123: 103898. doi：10.1016/j.compbiomed.2020.103898
[10]	Khan W, Zaki N, Ali L. Intelligent pneumonia identification from chest X-rays: a systematic literature review[J]. IEEE Access, 2021, 9: 51747-71. doi：10.1109/ACCESS.2021.3069937
[11]	李双双, 侯震, 刘娟, 等. 影像组学分析与建模工具综述[J]. 中国医学物理学杂志, 2018, 35(9): 1043-9. doi：10.3969/j.issn.1005-202X.2018.09.010
[12]	Wang K, An Y, Zhou JC, et al. A novel Multi-Level feature selection method for radiomics[J]. Alex Eng J, 2023, 66: 993-9. doi：10.1016/j.aej.2022.10.069
[13]	Horvat N, Papanikolaou N, Koh DM. Radiomics beyond the hype: a critical evaluation toward oncologic clinical use[J]. Radiol Artif Intell, 2024, 6(4): e230437. doi：10.1148/ryai.230437
[14]	Lee E, Cao LD, Vishwa P, et al. Nimg-08. an integrated informatics model combining clinical factors, radiomics and a novel connectomics framework to distinguish pathologically-proven radionecrosis from progression in treated brain metastases[J]. Neuro Oncol, 2022, 24(): vii163. doi：10.1093/neuonc/noac209.627
[15]	Ferretti M, Corino VDA. Integrating radiomic and 3D autoencoder-based features for Non-Small Cell Lung Cancer survival analysis[J]. Comput Methods Programs Biomed, 2025, 258: 108496. doi：10.1016/j.cmpb.2024.108496
[16]	Choi W, Oh JH, Riyahi S, et al. Radiomics analysis of pulmonary nodules in low-dose CT for early detection of lung cancer[J]. Med Phys, 2018, 45(4): 1537-49. doi：10.1002/mp.12820
[17]	Meng XC, Xia W, Xie PY, et al. Preoperative radiomic signature based on multiparametric magnetic resonance imaging for noninvasive evaluation of biological characteristics in rectal cancer[J]. Eur Radiol, 2019, 29(6): 3200-9. doi：10.1007/s00330-018-5763-x
[18]	Kong XY, He YM, Song ZH, et al. Deep probabilistic principal component analysis for process monitoring[J]. IEEE Trans Neural Netw Learn Syst, 2025, 36(4): 7422-36. doi：10.1109/tnnls.2024.3386890
[19]	Wang LW, Tao SY, Zhu P, et al. Data-driven topology optimization with multiclass microstructures using latent variable Gaussian process[J]. J Mech Des, 2021, 143(3): 031708. doi：10.1115/1.4048628
[20]	Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative adversarial networks[J]. Commun ACM, 2020, 63(11): 139-44. doi：10.1145/3422622
[21]	Ning ZY, Du DH, Tu C, et al. Relation-aware shared representation learning for cancer prognosis analysis with auxiliary clinical variables and incomplete multi-modality data[J]. IEEE Trans Med Imaging, 2022, 41(1): 186-98. doi：10.1109/tmi.2021.3108802
[22]	Ning ZY, Xiao Q, Feng QJ, et al. Relation-induced multi-modal shared representation learning for Alzheimer's disease diagnosis[J]. IEEE Trans Med Imaging, 2021, 40(6): 1632-45. doi：10.1109/tmi.2021.3063150
[23]	Ning ZY, Lin ZH, Xiao Q, et al. Multi-constraint latent representation learning for prognosis analysis using multi-modal data[J]. IEEE Trans Neural Netw Learn Syst, 2023, 34(7): 3737-50. doi：10.1109/tnnls.2021.3112194
[24]	Tahir AM, Chowdhury MEH, Qiblawey Y, et al. COVID-QU-Ex Dataset[DB/OL]. [2026-01-12]. .
[25]	van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational radiomics system to decode the radiographic phenotype[J]. Cancer Res, 2017, 77(21): e104-7. doi：10.1158/0008-5472.can-17-0339
[26]	Brown G, Pocock AC, Zhao MJ, et al. Conditional likelihood maximisation: a unifying framework for mutual information feature selection[J]. J Mach Learn Res, 2012, 13(1): 27-66. doi：10.1080/00207179.2012.669851
[27]	Hoque N, Bhattacharyya DK, Kalita JK. MIFS-ND: a mutual information-based feature selection method[J]. Expert Syst Appl, 2014, 41(14): 6371-85. doi：10.1016/j.eswa.2014.04.019
[28]	Wang R, Bian JT, Nie FP, et al. Unsupervised discriminative projection for feature selection[J]. IEEE Trans Knowl Data Eng, 2022, 34(2): 942-53. doi：10.1109/TKDE.2020.2983396
[29]	Kavitha KR, Neeradha K, Athira, et al. Laplacian score and top scoring pair feature selection algorithms[C]//2020 Fourth International Conference on Computing Methodologies and Communication (ICCMC). March 11-13, 2020. Erode, India. IEEE, 2020: 214-9. doi：10.1109/ICCMC48092.2020.ICCMC-00041
[30]	Wang YS, Yao HX, Zhao SC. Auto-encoder based dimensionality reduction[J]. Neurocomputing, 2016, 184: 232-42. doi：10.1016/j.neucom.2015.08.104
[31]	Bugata P, Drotar P. On some aspects of minimum redundancy maximum relevance feature selection[J]. Sci China Inf Sci, 2019, 63(1): 112103. doi：10.1007/s11432-019-2633-y
[32]	Yousaf M, Rehman TU, Jing L. An extended isomap approach for nonlinear dimension reduction[J]. SN Comput Sci, 2020, 1(3): 160. doi：10.1007/s42979-020-00179-y
[33]	Cieslak MC, Castelfranco AM, Roncalli V, et al. T-Distributed Stochastic Neighbor Embedding (t-SNE): a tool for eco-physio-logical transcriptomic analysis[J]. Mar Genom, 2020, 51: 100723. doi：10.1016/j.margen.2019.100723
[34]	Wang LJ, Mo TL, Wang XT, et al. A hierarchical fusion framework to integrate homogeneous and heterogeneous classifiers for medical decision-making[J]. Knowl Based Syst, 2021, 212: 106517. doi：10.1016/j.knosys.2020.106517
[35]	He Q, Li X, Nathan Kim DW, et al. Feasibility study of a multi-criteria decision-making based hierarchical model for multi-modality feature and multi-classifier fusion: Applications in medical prognosis prediction[J]. Inf Fusion, 2020, 55: 207-19. doi：10.1016/j.inffus.2019.09.001
[36]	Tayarani-Najaran MH. A novel ensemble machine learning and an evolutionary algorithm in modeling the COVID-19 epidemic and optimizing government policies[J]. IEEE Trans Syst Man Cybern Syst, 2022, 52(10): 6362-72. doi：10.1109/tsmc.2022.3143955
[37]	Nan Y, Del Ser J, Walsh S, et al. Data harmonisation for information fusion in digital healthcare: a state-of-the-art systematic review, meta-analysis and future research directions[J]. Inf Fusion, 2022, 82: 99-122. doi：10.1016/j.inffus.2022.01.001
[38]	Mattoli MV, Calcagni ML, Taralli S, et al. How often do we fail to classify the treatment response with [(18)F] FDG PET/CT acquired on different scanners? data from clinical oncological practice using an automatic tool for SUV harmonization[J]. Mol Imaging Biol, 2019, 21(6): 1210-9. doi：10.1007/s11307-019-01342-5
[39]	Reuzé S, Orlhac F, Chargari C, et al. Prediction of cervical cancer recurrence using textural features extracted from 18F-FDG PET images acquired with different scanners[J]. Oncotarget, 2017, 8(26): 43169-79. doi：10.18632/oncotarget.17856