The Idea: Because fastrad uses standard PyTorch tensor operations, we can theoretically preserve the computational graph (requires_grad=True).
Why it's cool: This would allow researchers to embed standard radiomics features inside a deep learning neural network. For example, you could train a CNN where the loss function directly optimizes for GLCM Contrast or Shape Sphericity, calculating backpropagation straight through the fastrad feature extractor!
The Idea: Because
fastraduses standard PyTorch tensor operations, we can theoretically preserve the computational graph (requires_grad=True).Why it's cool: This would allow researchers to embed standard radiomics features inside a deep learning neural network. For example, you could train a CNN where the loss function directly optimizes for GLCM Contrast or Shape Sphericity, calculating backpropagation straight through the
fastradfeature extractor!