Brief Review — ResHNet: Spectrograms Based Efficient Heart Sounds Classification Using Stacked Residual Networks

ResHNet, Using Residual Blocks in ResNet

Sik-Ho Tsang
2 min readJan 7, 2024

ResHNet: Spectrograms Based Efficient Heart Sounds Classification Using Stacked Residual Networks
ResHNet
, Institute for Infocomm Research, A*STAR; National University of Singapore; Chongqing Jiaotong University; YITU Tech
2019 BHI (Sik-Ho Tsang @ Medium)

Heart Sound Classification
20132023 [2LSTM+3FC, 3CONV+2FC] [NRC-Net] [Log-MelSpectrum+Modified VGGNet] [CNN+BiGRU]
==== My Other Paper Readings Are Also Over Here ====

  • Spectrogram is used as input to train the proposed residual network based classifier, namely ResHNet, for identifying normal and abnormal heart sounds.

Outline

  1. ResHNet
  2. Results

1. ResHNet

1.1. Spectrogram

Spectrogram

Heart sound is converted into spectrogram as CNN input.

1.2. ResHNet

A Residual Block

Residual Block in ResNet is utilized to constructed ResHNet.

ResHNet
  • ReLU is used.
  • Batch Normalization is incorporated in the residual module.
  • The initial convolution layer before reducing spatial dimensions will have a total of 64 filters. Then 3 sets of residual modules are stacked with three convolution layers in each module learning 32, 32 and 128 convolution filters respectively.
  • Then another four sets of residual modules are stacked with three convolution layers in each module learning 64, 64 and 256 filters respectively after reducing the spatial dimensions.
  • The same series of action is repeated for 6 sets of residual modules.
  • Average Pooling is performed after the final spatial dimension reduction and on top of it a softmax classifier is applied to get the output.

2. Results

SOTA Comparison

ResHNet obtains the highest Sensitivity, Specificity and Mean Accuracy (Score).

--

--

Sik-Ho Tsang

PhD, Researcher. I share what I learn. :) Linktree: https://linktr.ee/shtsang for Twitter, LinkedIn, etc.