eeg data and unet question

[christinazavou]

Hello, thanks for sharing the code and all your great work:)

I'm trying to replicate the work in "High Frequency EEG Artifact Detection with Uncertainty via Early Exit Paradigm" which is using the early-exit paradigm as well, and I run into some questions. Since this code is open sourced I'm asking them here so more people interested in it can follow.

• You are using TUH Artifact v2.0.0. In that version the labels are given per bipolar channel (e.g. the tcp_ar_01 montage has 22 channels). Which data have you used? The ones coming from each of the 22 channels from tcp_ar_01? Also, you mention the data are selected "based on an analysis of the duration of the majority of artifacts" ... could you please clarify what do you mean? I would think that you take for example 10 sec segments where artifact appeared for at least 1 sec .. did you do something similar?

• Regarding the implementation, in table 2 the exit block shows an Upsample followed by a Conv1D ... but in fig. 2 the exit block is represented by a Conv1d, an Activation and a second Conv1d. Which one is the correct?

Some minor implementation details (please correct me if I'm wrong):

- in table 2 it shows that you go from shape 5x2500 to shape 5x1250 with Max pooling of kernel size 2 and stride 1 .. 
however i believe this is achieved with kernel size 1 and stride 2 or kernel size 3, stride 2 and padding 1.
Additionally to go from 9x625 to 9x312 we need to use kernel size 3 stride 2 and padding 0.

- in table 2 I don't see how from input shape [1 channel, 2500 time steps] we can use 1d convolution with 
kernel 4 and stride 1 and get the output shape [5, 2500]. I can get that output shape using kernel 5 and 
padding 2. (The same for the rest of the convolutions in the encoder).

thanks in advance, christina

[lorenaqendro]

Hi @christinazavou, Thank you for your interest in the paper and your kind words.

Please have a look at the following comments which I hope answer your questions:

• Yes, that's right. We use the signals stemming from the edf_01_tcp_ar’s channels. What we mean by “based on an analysis of the duration of the majority of artifacts” is: We decided to consider a signal duration of 10s since by analyzing the duration of each artifact in the dataset, we saw that the majority of them last less than 10s. It’s mainly to justify the signal length decision. So your understanding is correct.
• For the implementation, please refer to Table 2.

• As you have described, you need the padding details to proceed. Our implementation is based on the temporal U-Net proposed by Perslev et al. where for the encoder and decoder convolutions they use the TensorFlow padding=“same”. For PyTorch it looks something like this:

  class Conv1d_padding_same(nn.Module):
  def __init__(self, in_channels, out_channels, kernel_size, dilation=1, stride=1):
      super().__init__()
      self.padding = math.ceil((1 - stride + dilation * (kernel_size - 1)) / 2)
      self.cut_last = (kernel_size % 2 == 0 and dilation % 2 == 1)
      self.conv = nn.Conv1d(in_channels, out_channels, kernel_size, padding=self.padding, dilation=dilation, stride=stride)
  
  def forward(self, x):
  
      if self.cut_last:
          return self.conv(x)[:, :, :-1]
      else:
          return self.conv(x)

  For the encoder/decoder Conv1d_padding_same, kernel_size=4 and stride=1 are used. Instead, for the exit and final output blocks, padding=0 and kernel_size=1 are used. This should hopefully solve the shaping issues.