心音异常高检出率模型MS-DTNet：融合多尺度卷积与双塔注意力的方法

doi:10.12122/j.issn.1673-4254.2026.06.24

摘要/Abstract

摘要：

目的提高心音异常的检出率，为自动化心音分析提供技术支持。方法本文提出了一种基于多尺度卷积与双塔注意力机制的模型（MS-DTNet），其通过深度可分离卷积以降低计算复杂度的同时结合多尺度卷积以并行获取不同感受野的特征，然后采用双塔结构兼顾短时与长时信息提取，并引入翻转注意力机制以进一步增强特征表征能力，从而提高对心音图分类的精度。结果使用公开数据集CinC2016对提出的模型进行五折交叉验证实验。在心音异常分类任务中，模型在测试集上的平均准确率、平均特异性、平均召回率、平均精确率、平均F1分数与平均ROC分别达到94.49%、93.44%、95.49%、93.81%、94.64%和98.65%，结果均优于其他几种模型。结论本研究提出的MS-DTNet模型在心音图分类表现出色，为后续的多源心音数据分析与临床辅助决策研究提供了有效的技术基础。

关键词: 心音图分类, 轻量化神经网络, 深度可分离卷积, 多尺度特征卷积, 注意力机制, 双塔机制

Abstract:

Objective To develop an efficient model for detecting abnormal heart sounds and providing technical support for automated phonocardiogram (PCG) analysis. Methods The proposed model, MS-DTNet, integrates multi-scale convolution and a dual-tower attention mechanism. The network employs depthwise separable convolution to achieve a lightweight architecture, applies multi-scale convolution to capture features across receptive fields in parallel, utilizes a dual-tower structure to extract both short- and long-term temporal information, and incorporates a flipped-attention mechanism to further enhance feature representation. Experiments were conducted on the public CinC2016 dataset using 5-fold cross-validation. Results In the abnormal heart sound classification task, the proposed model achieved an average accuracy of 94.49%, a specificity of 93.44%, a sensitivity of 95.49%, a precision of 93.81%, F1-score of 94.64%, and ROC-AUC of 98.65%, outperforming all the comparison models. Conclusion The proposed MS-DTNet model demonstrates excellent performance in PCG classification and provides technical assistance for multi-source heart sound analysis and clinical decision support.

Key words: phonocardiogram classification, lightweight neural network, depthwise separable convolution, multi-scale convolution, attention mechanism, dual-tower architecture

温耀棋, 卢学麒, 吴秋岑, 陈华元, 陈超敏. 心音异常高检出率模型MS-DTNet：融合多尺度卷积与双塔注意力的方法[J]. 南方医科大学学报, 2026, 46(6): 1444-1454.

Yaoqi WEN, Xueqi LU, Qiucen WU, Huayuan CHEN, Chaomin CHEN. MS-DTNet: an efficient model combining multi-scale convolution and dual-tower attention for detecting abnormal heart sounds[J]. Journal of Southern Medical University, 2026, 46(6): 1444-1454.

图/表 14

表1 CinC2016数据集统计信息

Tab.1 CinC2016 dataset statistics

Dataset	Label	Sum	Percentage	Describe
Training set and validation set	Normal	2575	79.5%	S1 and S2 clear
Training set and validation set	Abnormal	665	20.5%	Too much noise
Testing Set	Normal	150	49.8%
Testing Set	Abnormal	151	50.2%
Sum		3541	100%

图1 不合格的心音图示例

Fig.1 An example of substandard phonocardiogram.

图2 有价值的心音图示例

Fig.2 An example of useful phonocardiogram.

表2 数据清洗结果

Tab.2 Result of data-cleaning

Total	Removed_few_peaks	Removed_low_Q	Recovered	Final_samples	Used_thresh
3240	46	48	43	3189	0.25

表3 数据处理后数据库统计信息

Tab.3 Database statistics after data processing

Class	Training set	Validation set	Testing set	Sum	Training set (smote)
Normal	2284	257	150	2691	2284
Abnormal	586	62	151	799	2284
Sum	2870	319	301	3490	4568

图3 MS-DTNet整体架构

Fig.3 Overall architecture of the MS-DTNet model.

表4 MS-DTNet网络参数设置

Tab.4 MS-DTNet parameter settings

Module/block	Layer	Input	Output	Description of Feature Map Changes
Input	PCG single-channel signal	1		Input shape(B, 1, T)
Stem	Conv(7×1, stride=2)+BN+ReLU $→$ Conv(3×1)+BN+ReLU	1	48 (bc)	Output shape: (B, 48, T/2)
LMS	Conv1×1 reduction + parallel 3/5/7 convolutions + concat	48	144 (bc×3)	Multi-scale feature concatenation
L_Tower_1	Light Res Block (stride=2)	144	96 (bc×2)	Downsample×2, includes flip attention mechanism
L_Tower_2	Light Res Block (stride=1)	96	192 (bc×4)	Channel number ×2
L_Tower_3	Light Res Block (stride=1)	192	384 (bc×8)	Channel number ×2
R_Tower_1	ResNeXt1DBlock(stride=2)	144	96 (bc×2)	Downsample×2, includes flip attention mechanism
R_Tower_2	ResNeXt1DBlock(stride=1)	96	192 (bc×4)	Channel number ×2
R_Tower_3	ResNeXt1DBlock(stride=1)	192	384 (bc×8)	Channel number ×2
Dual-tower fusion	Element-wise addition	384	384	Feature alignment and summation
ReNeXtBlock	ResNeXt1DBlock(stride=1)	384	384	Channel number unchanged
Global pooling	AvgPool+MaxPool $→$ concat	384	768 (384+384)	Channel number ×2
Classifier	Linear(768 $→$ 384)+ReLU+Dropout $→$ Linear(384 $→$ 1)	768	1	Binary classification output

表4 MS-DTNet网络参数设置

Tab.4 MS-DTNet parameter settings

Module/block	Layer	Input	Output	Description of Feature Map Changes
Input	PCG single-channel signal	1		Input shape(B, 1, T)
Stem	Conv(7×1, stride=2)+BN+ReLU $→$ Conv(3×1)+BN+ReLU	1	48 (bc)	Output shape: (B, 48, T/2)
LMS	Conv1×1 reduction + parallel 3/5/7 convolutions + concat	48	144 (bc×3)	Multi-scale feature concatenation
L_Tower_1	Light Res Block (stride=2)	144	96 (bc×2)	Downsample×2, includes flip attention mechanism
L_Tower_2	Light Res Block (stride=1)	96	192 (bc×4)	Channel number ×2
L_Tower_3	Light Res Block (stride=1)	192	384 (bc×8)	Channel number ×2
R_Tower_1	ResNeXt1DBlock(stride=2)	144	96 (bc×2)	Downsample×2, includes flip attention mechanism
R_Tower_2	ResNeXt1DBlock(stride=1)	96	192 (bc×4)	Channel number ×2
R_Tower_3	ResNeXt1DBlock(stride=1)	192	384 (bc×8)	Channel number ×2
Dual-tower fusion	Element-wise addition	384	384	Feature alignment and summation
ReNeXtBlock	ResNeXt1DBlock(stride=1)	384	384	Channel number unchanged
Global pooling	AvgPool+MaxPool $→$ concat	384	768 (384+384)	Channel number ×2
Classifier	Linear(768 $→$ 384)+ReLU+Dropout $→$ Linear(384 $→$ 1)	768	1	Binary classification output

图4 轻量化多尺度感受野模块

Fig.4 Light multi-scale field module.

图5 轻量化残差块

Fig.5 Light Res Block.

图6 ResNeXt 模块

Fig.6 ResNeXt Block.

表5 实验配置

Tab.5 Experiment configuration

Item	Configuration/hyperparameter
Model	MS-DTNet (input channels=1)
Optimizer	AdamW
Initial learning rate	0.001
Learning rate scheduler	Cosine annealingLR (T_max=20)
Batch size	32
Epochs	40
Loss function	BCE withlogits loss
Random seed	42

表6 MS-DTNet的分类性能（五折交叉验证）

Tab.6 Classification performance of MS-DTNet （5-fold cross validation）

Fold	Acc (%)	Spe (%)	Recall (%)	Pre (%)	F1 (%)	AUC (%)
1	97.34	97.97	96.73	98.01	97.37	99.49
2	95.68	96.60	94.80	96.69	95.73	98.52
3	96.01	96.00	96.02	96.02	96.02	98.64
4	97.01	96.00	96.03	96.10	96.02	98.91
5	97.01	95.33	98.68	95.51	97.07	99.38
Average	96.61±0.72	96.38±0.99	96.45±1.42	96.47±0.96	96.65±0.72	98.99±0.43

表7 不同文献中分类算法的平均准确率对比结果

Tab.7 Comparison of classification avg-acc from this study with other methods

Author	Framework	Core Idea	Dataset	ACC (%)
Bao XQ, et al^[18]	BiLSTM	Temporal modeling: using bidirectional recurrent units to capture PCG sequence dependencies	CinC 2016	92.64%
Xiao B, et al^[19]	CNN	Classic convolutional network: end-to-end extraction of local temporal features	CinC 2016	93.00%
Orozco-Reyes L, et al^[20]	WST-based 1D-CNN	Fusion of multi-scale wavelet scattering features with CNN	CinC 2016	85.90%
Harimi A, et al^[21]	2DCNNs	2D convolutional classification framework using time-frequency images as input	CinC 2016	92.19%
Suchithra KP, et al^[22]	CNN-LSTM	Hybrid architecture combining convolutional and recurrent modules	CinC 2016	89.68%
Karhade J, et al^[46]	TF-Deep CNN	Deep convolutional modeling of time-frequency features	CinC 2016	85.16%
Morshed M, et al^[47]	1D-CNN	Lightweight 1D convolutional model for heart sound recognition	CinC 2016	92.59%
Our method	MS-DTNet	Multi-scale+Dual-Tower structure+Flip Attention mechanism	CinC 2016	96.61%

表8 消融实验结果

Tab.8 Ablation experimental results

Model	Acc (%)	Spe (%)	Recall (%)	Pre (%)	F1 (%)	AUC (%)
Baseline	81.95±2.14	71.87±8.52	89.93±7.49	76.73±3.41	82.88±2.22	89.28±2.82
Baseline-LMS	83.14±4.40	74.00±5.21	94.17±2.56	80.87±4.74	86.03±2.91	91.74±2.15
Baseline-LMS-RT	90.64±3.25	85.87±6.69	97.35±1.39	85.05±5.66	92.24±2.16	96.59±2.53
Baseline-LMS-LT	91.84±4.79	87.33±0.30	98.28±0.68	86.58±5.25	93.44±5.81	97.19±2.40
Baseline-DT	94.70±2.02	94.00±1.89	97.35±0.62	89.07±1.99	95.33±1.11	98.23±1.09
MS-DTNet-noflip	92.17±1.08	89.20±3.25	97.09±0.99	93.56±1.96	93.20±1.72	97.20±0.72
MS-DTNet-SE	93.10±0.77	90.53±6.25	97.62±1.95	92.28±1.44	94.22±1.79	97.59±1.90
MS-DTNet-CBAM	92.77±0.88	90.80±1.26	96.69±1.11	93.20±1.31	94.85±1.62	97.88±0.62
MS-DTNet	96.61±0.72	96.38±0.99	96.45±1.42	96.47±0.96	96.65±0.72	98.99±0.43

参考文献 49

[1]	World Health Organization. Global report on hypertension: the race against a silent killer[M]. World Health Organization, 2023.
[2]	刘明波, 何新叶, 杨晓红, 等. 《中国心血管健康与疾病报告2023》要点解读[J]. 中国心血管杂志, 2024, 29(4): 305-24. doi：10.3969/j.issn.1007-5410.2024.04.002
[3]	Rangayyan R M, Lehner R J. Phonocardiogram signal analysis: a review[J]. Critical Reviews Biomedical Engineering, 1987, 15(3): 211-36.
[4]	Krishnan PT, Balasubramanian P, Umapathy S. Automated heart sound classification system from unsegmented phonocardiogram (PCG) using deep neural network[J]. Phys Eng Sci Med, 2020, 43(2): 505-15. doi：10.1007/s13246-020-00851-w
[5]	LeCun Y, Bengio Y, Hinton G. Deep learning[J]. Nature, 2015, 521(7553): 436-44. doi：10.1038/nature14539
[6]	Chen W, Sun Q, Chen XM, et al. Deep learning methods for heart sounds classification: a systematic review[J]. Entropy, 2021, 23(6): 667. doi：10.3390/E23060667
[7]	Misal A, Sinha G R. Denoising of PCG signal by using wavelet transforms[J]. Advances in Computational Research, 2012, 4(1): 46-9. doi：10.26634/jcs.1.1.1740
[8]	Farzam B, Shirazi J. The diagnosis of heart diseases based on PCG signals using MFCC coefficients and SVM classifier[J]. IJISET-International J Innovative Science, Engineering & Technology, 2014, 1(10): 230-8.
[9]	Ari S, Hembram K, Saha G. Detection of cardiac abnormality from PCG signal using LMS based least square SVM classifier[J]. Expert Syst Appl, 2010, 37(12): 8019-26. doi：10.1016/j.eswa.2010.05.088
[10]	Ghosh SK, Tripathy RK, Ponnalagu R N. Classification of PCG signals using Fourier-based synchrosqueezing transform and support vector machine[C]//2021 IEEE Sensors. October 31-November 3, 2021. Sydney, Australia. IEEE, 2021: 1-4. DOI:10.1109/sensors47087.2021. 9639687 .
[11]	Arslan Ö. Automated detection of heart valve disorders with time-frequency and deep features on PCG signals[J]. Biomed Signal Process Control, 2022, 78: 103929. doi：10.1016/j.bspc.2022.103929
[12]	Zhang X, Lin MJ, Hong Y, et al. MSFT: a multi-scale feature-based transformer model for arrhythmia classification[J]. Biomed Signal Process Control, 2025, 100: 106968. doi：10.1016/j.bspc.2024.106968
[13]	Alzubaidi L, Zhang JL, Humaidi AJ, et al. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions[J]. J Big Data, 2021, 8(1): 53. doi：10.1186/s40537-021-00444-8
[14]	Sherstinsky A. Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network[J]. Phys D Nonlinear Phenom, 2020, 404: 132306. doi：10.1016/j.physd.2019.132306
[15]	Wang J, Yang Y, Mao JH, et al. CNN-RNN: a unified framework for multi-label image classification[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 27-30, 2016. Las Vegas, NV, USA. IEEE, 2016: 2285-94. DOI:10.1109/cvpr.2016.251 .
[16]	Han K, Xiao A, Wu E, et al. Transformer in transformer[J]. Advances Neural Information Processing Systems, 2021, 34: 15908-19.
[17]	Lin MJ, Zhang X, Hong Y, et al. RDiffGAN-ECG: High-quality specific R-peaks conditional ECG generation using denoising diffusion generative adversarial network with state space models[J]. Biomed Signal Process Control, 2026, 112: 108544. doi：10.1016/j.bspc.2025.108544
[18]	Bao XQ, Xu YJ, Kamavuako EN. The effect of signal duration on the classification of heart sounds: a deep learning approach[J]. Sensors, 2022, 22(6): 2261. doi：10.3390/s22062261
[19]	Xiao B, Xu YQ, Bi XL, et al. Heart sounds classification using a novel 1-D convolutional neural network with extremely low parameter consumption[J]. Neurocomputing, 2020, 392: 153-9. doi：10.1016/j.neucom.2018.09.101
[20]	Orozco-Reyes L, Alonso-Arévalo MA, García-Canseco E, et al. A deep-learning approach to heart sound classification based on combined time-frequency representations[J]. Technologies, 2025, 13(4): 147. doi：10.3390/technologies13040147
[21]	Harimi A, Majd Y, Gharahbagh AA, et al. Classification of heart sounds using chaogram transform and deep convolutional neural network transfer learning[J]. Sensors, 2022, 22(24): 9569. doi：10.3390/s22249569
[22]	Suchithra KP, Mohan N. Deep learning-based classification ofPCG signals using chaogram transform andCNN-LSTM network[C]//Innovative Computing and Communications. Singapore: Springer, 2024: 567-76. doi：10.1007/978-981-97-3817-5_40
[23]	Zhu LX, Qiu WY, Ma Y, et al. LEPCNet: a lightweight end-to-end PCG classification neural network model for wearable devices[J]. IEEE Trans Instrum Meas, 2024, 73: 1-11. doi：10.1109/tim.2023.3315401
[24]	Barnawi A, Boulares M, Somai R. Simple and powerful PCG classification method based on selection and transfer learning for precision medicine application[J]. Bioengineering, 2023, 10(3): 294. doi：10.3390/bioengineering10030294
[25]	Touahria R, Hacine-Gharbi A, Ravier P, et al. Relevant multi domain features selection based on mutual information for heart sound classification[C]//Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods. February 24-26, 2024. Rome, Italy. SCITEPRESS-Science and Technology Publications, 2024: 918-23. DOI:10.5220/0012565300003654 .
[26]	Sondermann M, Bisgin P, Tschorn N, et al. Comparative Analysis of CNN and Transformer Architectures with Heart Cycle Normalization for Automated Phonocardiogram Classification[J]. arXiv preprint arXiv:, 2025. doi：10.1109/EMBC58623.2025.11251770
[27]	Liu CY, Springer D, Li Q, et al. An open access database for the evaluation of heart sound algorithms[J]. Physiol Meas, 2016, 37(12): 2181-213. doi：10.1088/0967-3334/37/12/2181
[28]	Egger J. PCG-cut: graph driven segmentation of the prostate central gland[J]. PLoS One, 2013, 8(10): e76645. doi：10.1371/journal.pone.0076645
[29]	Cheadle C, Vawter MP, Freed WJ, et al. Analysis of microarray data using Z score transformation[J]. J Mol Diagn, 2003, 5(2): 73-81. doi：10.1016/s1525-1578(10)60455-2
[30]	Sharma S. Digital signal processing[M]. Khanna Publishing, 2025. doi：10.1016/j.bspc.2024.106826
[31]	Groeneveld RA, Meeden G. Measuring skewness and kurtosis[J]. Statistician, 1984, 33(4): 391. doi：10.2307/2987742
[32]	KaySteven M. Fundamentals of statistical signal processing: estimation theory[M]. Technometrics,1995. doi：10.2307/1269750
[33]	Boashash B.Time-Frequency Signal Analysis and Processing: A Comprehensive Review[M]. Oxford: Academic Press. doi：10.1016/b978-0-12-398499-9.09989-1
[34]	Ramı́rez J, Segura JC, Benı́tez C, et al. Efficient voice activity detection algorithms using long-term speech information[J]. Speech Commun, 2004, 42(3/4): 271-87. doi：10.1016/j.specom.2003.10.002
[35]	Rabiner L. On the use of autocorrelation analysis for pitch detection[J]. IEEE Trans Acoust, Speech, Signal Process, 1977, 25(1): 24-33. doi：10.1109/tassp.1977.1162905
[36]	Kedem B. Spectral analysis and discrimination by zero-crossings[J]. Proc IEEE, 1986, 74(11): 1477-93. doi：10.1109/proc.1986.13663
[37]	Sujit NR, Kumar CS, Rajesh CB. Improving the performance of cardiac abnormality detection from PCG signal[J]. AIP Conf Proc, 2016, 1715: 020053. doi：10.1063/1.4942735
[38]	Elizar E, Zulkifley MA, Muharar R, et al. A review on multiscale-deep-learning applications[J]. Sensors, 2022, 22(19): 7384. doi：10.3390/s22197384
[39]	He Q, Li XK, Cai B. Graph neural network recommendation algorithm based on improved dual tower model[J]. Sci Rep, 2024, 14: 3853. doi：10.1038/s41598-024-54376-3
[40]	Meng T, Tao Y, Chen ZQ, et al. Depth evaluation for metal surface defects by eddy current testing using deep residual convolutional neural networks[J]. IEEE Trans Instrum Meas, 2021, 70: 1-13. doi：10.48550/arXiv.2104.02472
[41]	Chollet F. Xception: deep learning with depthwise separable convolutions[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). July 21-26, 2017. Honolulu, HI. IEEE, 2017: 1251-8. DOI:10.1109/cvpr. 2017. 195 .
[42]	Hu J, Shen L, Sun G. Squeeze-and-excitation networks[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. June 18-23, 2018. Salt Lake City, UT. IEEE, 2018: 7132-41. DOI:10.1109/cvpr.2018.00745 .
[43]	Cohen T, Welling M. Group equivariant convolutional networks[C]//International conference on machine learning. PMLR, 2016: 2990-9. doi：10.48550/arXiv.1602.07576
[44]	Liu B, Wang C, Liu X, et al. HyMamba: Mamba with Hybrid Geometry-Feature Coupling for Efficient Point Cloud Classification[J]. arXiv preprint arXiv:, 2025.
[45]	Rastogi D, Johri P, Tiwari V, et al. Multi-class classification of brain tumour magnetic resonance images using multi-branch network with inception block and five-fold cross validation deep learning framework[J]. Biomed Signal Process Control, 2024, 88: 105602. doi：10.1016/j.bspc.2023.105602
[46]	Karhade J, Dash S, Ghosh SK, et al. Time-frequency-domain deep learning framework for the automated detection of heart valve disorders using PCG signals[J]. IEEE Trans Instrum Meas, 2022, 71: 1-11. doi：10.1109/tim.2022.3163156
[47]	Morshed M, Fattah SA. A deep neural network for heart valve defect classification from synchronously recorded ECG and PCG[J]. IEEE Sens Lett, 2023, 7(9): 1-4. doi：10.1109/lsens.2023.3307053
[48]	Woo S, Park J, Lee JY, et al. CBAM: convolutional block attention module[C]//Computer Vision-ECCV 2018. Cham: Springer, 2018: 3-19. doi：10.1007/978-3-030-01234-2_1
[49]	Tian GY, Lian C, Xu BR, et al. Classification of phonocardiogram based on multi-view deep network[J]. Neural Process Lett, 2023, 55(4): 3655-70. doi：10.1007/s11063-022-10771-3