Reduction of Acquisition Time in FTIR Spectroscopy via Spectral Super-Resolution by Deep Learning.

Area of Science:

• Spectroscopy
• Medical Imaging
• Computational Biology

Background:

• Fourier transform infrared spectroscopy (FTIR) is vital for molecular analysis in clinical diagnosis.
• Enhancing spectral resolution in FTIR is critical for detailed molecular insights.
• Current limitations include lengthy acquisition times and the need for high spectral resolution.

Purpose of the Study:

• To develop and evaluate deep-learning-based spectral super-resolution models for FTIR.
• To reconstruct high-resolution IR spectral images from low-resolution data.
• To reduce FTIR acquisition time without compromising spectral quality.

Main Methods:

• Development of spectral super-resolution models using residual networks and U-Net (SSR-ResUNet).
• Incorporation of 1D, 2D, and 3D Convolutional Neural Networks (CNNs).
• Training and testing on real FTIR images from renal graft tissue sections.

Main Results:

• SSR-ResUNet models achieved high performance (low RMSE, high SSIM).
• Deep learning models outperformed traditional interpolation methods.
• Spatial histological structures were preserved, comparable to K-means clustering.
• Acquisition time was reduced by up to 87.5%.

Conclusions:

• The developed spectral super-resolution approach significantly enhances FTIR spectral resolution.
• This method offers an efficient solution for reducing IR image acquisition time.
• Advancements enable faster spectral imaging with preserved molecular information, supporting clinical applications.