TL;DR
This paper tackles the challenge of accurately diagnosing multiple types of heart diseases from phonocardiogram (PCG) recordings, which are often noisy and lack large, diverse datasets. They propose a method that converts 1D PCG signals into 3D time-frequency spectrograms and then leverages transfer learning with ten different deep neural network architectures to classify six categories of heart sounds, achieving high accuracy.
What Problem Does This Paper Address?
Auscultation, listening to heart sounds with a stethoscope, is a fundamental but subjective method for diagnosing cardiovascular diseases. While effective, a doctor's skill can vary, leading to potential misdiagnoses, especially in developing regions. Existing automated PCG analysis methods often struggle with noisy signals, limited datasets for specific conditions, and accurately distinguishing between multiple disease categories. This paper aims to bridge this gap by developing a robust and accurate computational tool for multi-class heart sound classification, specifically addressing the need for better diagnostic aids.
Methods
The core of this research lies in a multi-step approach:
- Data Fusion: They combined an existing public database (Yaseen) with a newly collected dataset of pulmonary hypertension (PH) recordings, creating a more comprehensive dataset with six distinct heart sound categories: Normal (N), Aortic Stenosis (AS), Mitral Regurgitation (MR), Mitral Stenosis (MS), Mitral Valve Prolapse (MVP), and Pulmonary Hypertension (PH).
- Signal Preprocessing: To handle variations in sampling rates and noise, signals were resampled to 2000 Hz, normalized, and filtered using a third-order Butterworth filter.
- Data Augmentation: A crucial step to improve model robustness and generalize to noisy environments. They employed a 1D signal augmentation technique that adds a "random_signal" with a "delta" parameter to the "original_signal," effectively simulating background noise variations. This increased the dataset size significantly.
- Time-Frequency Representation: The 1D PCG signals were transformed into 3D time-frequency images (spectrograms) using the Continuous Wavelet Transform (CWT) with Morse analysis. This conversion is believed to enhance feature extraction capabilities for deep learning models.
- Transfer Learning: Ten pre-trained deep learning models (including GoogleNet, ResNet101, DenseNet201, DarkNet19, etc.) were fine-tuned for the PCG classification task. This approach leverages knowledge learned from large image datasets to improve performance on the smaller, specialized heart sound dataset.
- 10-Fold Cross-Validation: Rigorous evaluation was performed using 10-fold cross-validation to ensure the model's performance was consistent across different data splits.
Datasets
The study utilized two datasets:
- Database A (Yaseen): A public database containing 1000 audio recordings across five categories: Normal (N), Mitral Stenosis (MS), Mitral Valve Prolapse (MVP), Mitral Regurgitation (MR), and Aortic Stenosis (AS). Each category had 200 recordings.
- Database B (Pulmonary Hypertension): A newly collected dataset from the Second Hospital of Dalian Medical University, containing 74 PH subjects with 102 recordings. For consistency, 200 recordings were selected.
After data augmentation, the combined dataset comprised 2400 PCG recordings, with 400 recordings per class.
Key Insights and Ideas
- Spectrograms as Rich Features: Converting 1D PCG signals into 3D time-frequency spectrograms proved to be a powerful representation, allowing deep learning models to extract more discriminative features than raw signals.
- Data Augmentation is Key for Robustness: The proposed 1D signal augmentation technique significantly improved the models' ability to classify heart sounds even in the presence of noise, a common issue in real-world PCG data.
- Transfer Learning Outperforms Simpler Models: Transfer learning models consistently outperformed simpler architectures like Bi-LSTM and basic CNNs when applied to spectrograms, highlighting the benefit of leveraging pre-trained knowledge.
- No Single "Best" Architecture: While several transfer learning models performed exceptionally well, the paper suggests that different architectures might be optimal depending on specific dataset characteristics or computational constraints.
Key Results
The study achieved impressive results:
- High Accuracy: Four transfer learning models (ResNet101, DenseNet201, DarkNet19, and GoogleNet) achieved an average accuracy of 0.98 for classifying the six categories of heart sound signals.
- Robustness: The models demonstrated high performance even with noisy heart sound recordings, thanks to the data augmentation strategy.
- Comparison: The proposed transfer learning approach significantly outperformed simpler methods like Bi-LSTM (21.67% accuracy) and a basic CNN on spectrograms (76.67% accuracy).
Limitations and Future Work
- Database Limitations: While the authors combined datasets, they acknowledge the general challenge of limited, diverse clinical databases for heart sound classification. The PH dataset, though new, is still relatively small.
- Computational Cost: Training and fine-tuning multiple deep learning models can be computationally intensive.
- Future Work: The authors suggest that their approach could be further refined by exploring more advanced data augmentation techniques and investigating other transfer learning architectures. They also highlight the potential for this technology to be deployed in remote areas or community hospitals for early disease detection.
Summary updated