Usefulness of Data Simulation for Training Deep Learning Denoising Algorithms in Infrared Spectral Histology.

Area of Science:

• Spectroscopy
• Medical Imaging
• Artificial Intelligence

Background:

• Fourier-transform infrared (FTIR) spectroscopy is crucial for analyzing tissue samples.
• Noise in FTIR spectral images complicates analysis and limits acquisition speed.
• Reducing scans per pixel to save time introduces significant noise challenges.

Purpose of the Study:

• To develop and evaluate a method for training deep learning models using simulated data for denoising FTIR spectral images of paraffin-embedded tissues.
• To improve the generalization and robustness of denoising models by incorporating diverse simulated spectral characteristics and noise types.

Main Methods:

• A simulated linear generative model was developed, incorporating Voigt, Gaussian, and Lorentzian band shapes and additive/multiplicative Gaussian and Poisson noise.
• Multiple simulated datasets were created to mimic real FTIR spectral data from paraffin-embedded tissues.
• A ResUNet-1D-CNN architecture was trained using simulated data, real data, and a combination of both.

Main Results:

• The performance of denoising models was significantly influenced by the specific configurations of the simulated spectra.
• Models trained on simulated data achieved performance comparable to those trained on real data.
• A combination of real and simulated data provided a slight performance advantage over simulated data alone.

Conclusions:

• Data simulation is a powerful tool for developing robust denoising techniques for spectral imaging.
• Training deep learning models with simulated data reduces the need for extensive real data acquisition, facilitating faster clinical deployment.
• This approach enables efficient development of denoising methods for FTIR spectral imaging in clinical settings.