Brief Review — Short-segment Heart Sound Classification Using an Ensemble of Deep Convolutional Neural Networks
TF-ECNN, Score-Level Fusion of 1D-CNN and 2D-CNN
Short-segment Heart Sound Classification Using an Ensemble of Deep Convolutional Neural Networks
TF-ECNN, by Universiti Teknologi Malaysia, and King Abdullah University of Science and Technology
2019 ICASSP, Over 90 Citations (Sik-Ho Tsang @ Medium)Heart Sound Classification
2013 … 2023 [2LSTM+3FC, 3CONV+2FC] [NRC-Net] [Log-MelSpectrum+Modified VGGNet] [CNN+BiGRU] [CWT+MFCC+DWT+CNN+MLP]
==== My Other Paper Readings Are Also Over Here ====
- A time-frequency CNN ensemble (TF-ECNN) is proposed, which combines the 1D-CNN and 2D-CNN based on score-level fusion of the class probabilities.
Outline
- TF-ECNN
- Results
1. TF-ECNN
1.1. 1D-CNN and 2D-CNN
The first CNN (1D-CNN) accepts one-dimensional PCG time series data as input (i.e., the raw heartbeat signal).
The second CNN (2D-CNN) uses the two-dimensional time-frequency feature maps of MFCCs and time-varying autoregressive (TV-AR) coefficients as input.
- The same network architecture is used for both networks.
1.2. TF-ECNN
In the TF-ECNN, both the 1D-CNN and 2DCNN are combined based on score-level fusion by summing over the outputs of softmax layers from two individual CNNs to produce fused class prediction probabilities.
2. Results
The proposed CNN models generally outperform the baseline classifiers considerably in most of the performance measures.
In particular, the 2D-CNN with MFCCs achieved the best performance in specificity and MAcc.
The TF-ECNN gives the highest accuracy and the second highest in sensitivity.
