Using Shapley additive explanations and saliency for interpretation of multiclass classification of PDX spectral data with deep neural networks

The identification of spectral markers differentiating disease states when using spectral data is challenging in the context of modelling with deep neural networks, particularly in scenarios where classification models are developed with multiple classes. While a number of approaches do exist which can provide an insight into the features which are learnt by deep learning models, in biophotonics and chemical imaging these have received relatively little attention. In the present work we pilot the use of Fourier Transform chemical imaging with two deep-learning interpretation approaches within the context of a multi-class classification problem. Fully connected neural networks are developed on unfolded chemical imaging data captured on patient-derived xenografts developed from a colorectal cancer model. Separately, Shapley additive explanations and saliency approaches are used to derive feature sets which are discriminatory for class within this experimental model of colorectal cancer. Preliminary results suggest that Shapley additive explanations provide differentiating spectral sets which may not be derived with saliency, although the feature sets which are identified are dependent upon spectral pretreatment methodology. A dual approach which employs both strategies may be an effective strategy for the identification of feature sets in this context.