A Suitable and Interpretable Methodology for FTIR Spectral Classification

We propose a suitable and interpretable methodology for FTIR spectral classification when using weakly-labelled data. A multi-scale CNN is implemented with first layer kernel widths chosen to match common FTIR peak widths, before an ensemble of these models (EMSCNN) is constructed using validation set voting. In the context of cancerous tissue classification from FTIR spectra, EMSCNN achieves a weighted mean of per-class accuracies of $83\pm6$%, beating all other models tested. A new semi-supervised VAE version of the CRIME framework is implemented to interpret the model, elucidating distinct pathways to each spectral classification. Multiple VAE architectures are investigated using convolutional or transformer encoders and linear or transformer decoders. Finally, linear weighted cosine similarity models are constructed using the VAE latent space and achieve similar performance to direct classification methods. Our code is available at https://anonymous.4open.science/r/colon_data_analysis-neurips.