Reduction of Acquisition Time in Fourier Transform Infrared Spectral Imaging by Deep Learning for Clinical Applications

In infrared Fourier transform spectral imaging applied to biomedical challenges, data quality is of primary importance to achieving clinical objectives. However, different noise sources affect the infrared signal coming from the sample. Generally, the number of scans per pixel is fixed to a high value in order to ensure a high signal-to-noise ratio. However, the higher the number of scans, the higher the acquisition time, which may be incompatible with clinical practice. The objective of this work is therefore to use deep learning techniques to efficiently reconstruct high-quality infrared images from poor-quality ones due to the short acquisition time on formalin-fixed paraffin-embedded tissue sections coming from renal graft recipients. From paired 1-scan (acquisition time of 0.062 s per pixel with a Spotlight 400, PerkinElmer) and 64-scan (acquisition time of 4 s per pixel with a Spotlight 400, PerkinElmer) infrared images, two deep learning architectures (autoencoder and ResUNet) and three different layer types (multilayer perceptron, 1D-CNN and 2D-CNN) were evaluated for different preprocessing steps of the 64-scan reference images. Results demonstrate that the combined application of atmospheric correction and EMSC preprocessing had a significant impact on the denoising performance of the models. Furthermore, ResUNet architecture combined with 1D-CNN is able to reconstruct high-quality infrared images from poor ones with high fidelity while saving over 95% of the acquisition time. Additional experiments show that from a histopathological point of view, the reconstructed images are approximately equivalent to 16-scan images. This work thus makes short acquisition time of infrared images compatible with high-quality data and a clinical routine.