PCG_FTSeg: Enhancement of Phonocardiogram Segmentation using Convolutional Neural Networks with Fourier transform Module

💡 Highlights

• Introduce an enhanced U-Net architecture integrated with Convolutional Fourier Transform (CF) modules that fuse time-domain convolution and frequency-domain analysis via FFT and iFFT, enabling precise extraction of both temporal and spectral features for accurate PCG segmentation.
• Employ a dual CF module strategy—applied consecutively in both encoder and decoder layers—to robustly differentiate between S1 and S2 heart sounds and background noise, thereby significantly improving segmentation performance even in the presence of heart murmurs.
• Validate the proposed approach on multiple datasets (internal PhysioNet 2016 and external PhysioNet 2022 and AMC), demonstrating superior performance with an average F1 score of 97.64% for S1 and S2 segmentation compared to state-of-the-art methods such as LR-HSMM, LSTM, and BiLSTM.

📄 Paper

This repository provides the official implementation code for the paper: "Enhancement of Phonocardiogram Segmentation using Convolutional Neural Networks with Fourier transform Module."

Authors: Changhyun Park, Keewon Shin, Jinew Seo, Hyunseok Lim, Gyeong Hoon Kim, Woo- Young Seo, Sung-Hoon Kim, Namkug Kim

MI2RL LAB

💾 Requirements

• numpy >= 1.21.0
• scipy >= 1.7.0
• torch >= 1.10.0
• pytorch-lightning >= 1.6.0
• monai >= 0.9.0
• scikit-learn >= 1.0.0
• neurokit2 >= 0.1.9
• livelossplot >= 0.5.5
• matplotlib >= 3.4.0
• librosa >= 0.10.2
• natsort >= 8.4.0

🧱 Repository Structure

├── config.py 
├── dataset.py 
├── main.py 
├── model.py 
├── modules.py 
├── utils.py 
├── image 
│ └── model_figure.png 
├── preprocessing 
│ └── signal_2016_and_AMC.py 
│ └── signal_2022.py

🗂️ Datasets

• Internal Dataset (Train/Validation):
  • Uses the publicly available PhysioNet 2016 dataset
• External Datasets (Test):
  • Public PhysioNet 2022 dataset
  • Asan Medical Center (AMC) data; however, due to security regulations in Korea, the AMC data cannot be disclosed

📜 Script Example

python main.py 
    --gpu 0 
    --ver ['appropriate version identifier each time you run the script'] 
    --toler 40 
    --featureLength 6144 
    --target_sr 1000 
    --lowpass 20_200 
    --year 2016 
    --fft 
    --twice

🎯 Results:

Comparison of Methods on MITHSDB Benchmark (Tolerance Window = 40 ms)

Internal (PhysioNet 2016)

Method	Se (%)	P+ (%)	F1 S1 (%)	F1 S2 (%)	F1 (%)
LR-HSMM	95.6	94.4	-	-	95.0
LSTM	93.50	92.47	94.27	91.67	92.95
BiLSTM	94.87	94.53	95.58	93.86	94.93
U-Net	91.72±0.10*	92.06±0.14*	94.72±0.11*	88.82±0.06*	91.77±0.02*
U-Net + CFT (Ours)	97.67±0.03	97.64±0.15	98.82±0.16	96.46±0.03	97.64±0.09

External - 1 (PhysioNet 2022)

Method	Se (%)	P+ (%)	F1 S1 (%)	F1 S2 (%)	F1 (%)
U-Net	89.74±0.27*	91.94±0.20*	90.81±0.47*	90.54±0.00*	90.67±0.24*
U-Net + CFT (Ours)	94.62±0.26	95.57±0.01	94.97±0.14	95.11±0.12	95.04±0.13

External - 2 (AMC)

Method	Se (%)	P+ (%)	F1 S1 (%)	F1 S2 (%)	F1 (%)
U-Net	97.83±0.30*	98.33±0.38*	98.04±0.42*	98.07±0.27*	98.06±0.35*
U-Net + CFT (Ours)	99.18±0.13	99.16±0.14	99.12±0.07	99.21±0.02	99.16±0.14

Note: Data are shown as mean ± standard deviation.
* p-value < 0.05 (paired t-test), comparing the proposed model vs. the other methods.
The best results are highlighted in bold.

• BiLSTM: Bidirectional long short-term memory
• CF: Convolutional Fourier transform
• LR-HSMM: Logistic regression-hidden semi-Markov model
• LSTM: Long short-term memory
• MITHSDB: Massachusetts Institute of Technology Heart Sound Database

📝 Citation

If you use this code or find it useful in your research, please cite our paper:

@article{pcg_ftseg,
  title={Enhancement of Phonocardiogram Segmentation using Convolutional Neural Networks with Fourier transform Module},
  author={Changhyun Park and Keewon Shin and Jinew Seo and Hyunseok Lim and Gyeong Hoon Kim and Woo-Young Seo and Sung-Hoon Kim and Namkug Kim},
  year={2025}
}

🙋🏻‍♂️ Acknowledgements

We would like to thank Junseong Lee (@junjslee) for his notable contributions to refactoring the code structure and setting up this GitHub repository.