Abstract
Convolutional neural networks (CNNs) and support vector machines (SVMs) have seen numerous applications within Raman spectroscopy. However, the age-old question remains: Which is better? To shine some light on the matter, the stability of the two machine learning techniques was probed by intentionally introducing spectral artefacts to transmission low wavenumber Raman spectroscopic data. The data consisted of synthetic microcalcifications buried under various depths of chicken breast. We found that an SVM yielded the best model with an area under curve (AUC) of 0.989 compared to 0.979 for the CNN. However, generally, SVMs were more susceptible to the spectral artefacts than CNNs. Additionally, the performance of CNNs and SVMs was not dependent on the magnitude of the shifts and stretches. An example is the linear stretches, where the AUC remained at 0.977 and 0.969 for both 2 and 5 cm−1 shifts for the CNN and SVM models, respectively.