I’m currently studying the solutions of the RSNA cervical spine fracture detection challenge on kaggle and I was wondering:
Are the concepts of feature extraction and embeddings related?
What many people did was training a CNN on the CT scans slice wise and only use the last activations before the classification layer as their output. This gives a vector of size n for every slice of the CT scan. This is again fed into a RNN or LSTM model for CT scan wise classification. So the output of the CNN is a tensor of n_slices * n_features (ignoring the batch size).
Isn’t this feature vector of size n basically an embedding of the CT scan slices?
I hope it’s clear what I mean by that 