Brief Review — Noise Robust Cardio Net (NRC-Net)
NRC-Net: Automated noise robust cardio net for detecting valvular cardiac diseases using optimum transformation method with heart sound signals
NRC-Net, by Bangladesh University of Engineering and Technology; Kookmin University; Cogninet Australia; University of Southern Queensland; University of Technology Sydney; Australian International Institute of Higher Education; University of New England;University of Southern Queensland; Taylor’s University; SRM Institute of Science and Technology; and Kumamoto University
2023 Elsevier J. BSPC (Sik-Ho Tsang @ Medium)
Heart Sound Classification
2023 [2LSTM+3FC & 3CONV+2FC]
==== My Other Paper Readings Are Also Over Here ====
- First, different spectrogram inputs, Mel Frequency Cepstral Coefficients (MFCCs), Short-Time Fourier Transform (STFT), Constant-Q Nonstationary Gabor Transform (CQT), and Continuous Wavelet Transform (CWT) has been used with VGG-16 to investigate which one are the best.
- Furthermore, Noise Robust Cardio Net (NRC-Net) is proposed, which is a lightweight model for heart sound classification using PCG signals contaminated with respiratory and random noises.
- Noisy Heart Sound Signals & Spectrogram Features
- Noise Robust Cardio Net (NRC-Net)
1. Noisy Heart Sound Signals & Spectrogram Features
1.1. Summary of Prior Works
Many insights can be observed here but one of the things is that different prior arts use different input features (2nd column) and different classifier categories (3rd Column).
- First, raw heart sound are mixed with AWGN or lung sound to obtain noisy heart sound.
- CQT, CWT, STFT and MFCC are used to transform heart sound into spectrograms.
- These spectrograms are fed into the proposed network, NRC-Net for 5-class heart sound classification (N, AS, MR, MS, MVP).
2. Noise Robust Cardio Net (NRC-Net)
2.1. Spatial Feature Extractor Block (SFEB)
- The 3-channel 224 × 224 images are fed into the input layer.
- SFEB is implemented with 6 consecutive convolutional layers.
- A Batch Normalization layer and ReLU layer are used after a convolutional layer.
2.2. Holistic Attention Block (HAB)
- HAB is inspired from the fire module in SqueezeNet.
- The fundamental component of HAB comprises a squeeze layer with a 1×1 convolution layer and an expanded layer that incorporates parallel convolution layers of sizes 1×1 and 3×3, followed by two successive convolutional layers.
- A concatenation operation merges these two convolution layers.
2.3. Temporal Feature Extractor Block (TFEB)
- The TFEB block is formed using two parallel LSTM layers concatenated together.
2.4. Terminal Classification Block (TCB)
- The extracted feature vector from TFFB is flattened.
- The TCB is formed using the 5 fully connected layer and 1 softmax layer fed by the TFEB.
- The probability nodes for each class are obtained at output.
- Dropout is employed following each fully connected layer.
3.1. SOTA Comparisons
When evaluated on noise-free data, the proposed network achieved an accuracy of 99.70%, surpassing the accuracies of [38,41] by 1.22% and 0.10%, respectively. Furthermore, the proposed model demonstrated a sensitivity of 99.58% and a specificity of 99.66%, outperforming  by 1.6% and 1.2%, respectively.
The proposed model achieved an overall accuracy of 97.40% when dealing with heart sounds corrupted with lung sounds, outperforming the work of  by 2.62%.
3.2. Different Transformations, Different Noise Levels, Different Models
CWT obtains the best performance among all transformations.
The proposed model demonstrated superior performance in all scenarios.
3.3. Model Size and Inference Time
The proposed NRC-Net obtains the smallest model size and shortest inference time.
- (There are many results shown in the paper, please kindly read the paper directly for more details.)