Influence of denoising on classification results in the context of hyperspectral data: High Definition FT-IR imaging.

Owing to the high information content about the biochemical composition of the sample, the implementation of Fourier-Transform Infrared Spectroscopy (FT-IR) in the clinic is currently under investigation by many researchers. Cancer biology with the use of histopathological models is one of the most explored application areas. Most of the publications show sensitivity of the method to be above 90%, however, it is still often not enough for clinical standards. Robust denoising techniques with an optimized classification model allow to shorten the experimental acquisition times which still are a bottleneck for FT-IR translation into the clinic. The main premise of this work is to evaluate denoising impact on classification results using spectral techniques: Savitzky Golay (SG), Wavelets (WV), Principal Component Analysis (PCA) and Minimum Noise Fraction (MNF); and spatial techniques: Deep Neural Network (DNN), Median Filter. Using denoising methods, especially MNF and PCA, gave significant improvement of the classification and prediction results. Moreover, the increase in pixel level accuracy for High Definition data (1.1 μm projected pixel size) was found to be dependent on the complexity of the histopathological class and reached even 43-44% level, while core level increase reached around 28%. Moreover, we investigated the impact of denoising methods on the spectral input to better understand the mechanism of such large improvement. The results presented here highlight the benefits and the importance of proper denoising for classification purposes of FT-IR imaging data.