Denoising and Baseline Correction of Low-Scan FTIR Spectra: A Benchmark of Deep Learning Models Against Traditional...

Area of Science:

• Spectroscopy
• Medical Imaging
• Artificial Intelligence

Background:

• High-quality Fourier Transform Infrared (FTIR) imaging requires extensive signal averaging, limiting clinical speed.
• Deep learning offers potential for accelerating FTIR imaging via single-scan reconstruction.
• Separating noise and baseline drift in FTIR spectra is an ill-posed problem, challenging for standard deep learning models.

Purpose of the Study:

• To develop a novel deep learning architecture for accelerated, high-quality FTIR imaging.
• To address the limitations of standard deep learning models in handling spectral noise and baseline drift.
• To improve the speed and accuracy of FTIR spectral analysis for clinical applications.

Main Methods:

• A physics-informed cascade Unet architecture was proposed, incorporating a deterministic Physics Bridge.
• The model separates denoising and baseline correction tasks, using an embedded SNIP layer for spectroscopic constraints.
• Performance was benchmarked against a single Unet and traditional Savitzky-Golay smoothing methods using human hypopharyngeal carcinoma cell data.

Main Results:

• The cascade Unet achieved a 51.3% reduction in Root Mean Square Error (RMSE) compared to raw single-scan inputs.
• It outperformed a single Unet (40.2% RMSE reduction) and traditional methods (33.7% RMSE reduction).
• The model eliminated spectral hallucinations and preserved peak intensity with higher fidelity than traditional smoothing.

Conclusions:

• The proposed cascade Unet is a robust solution for diagnostic-grade FTIR imaging.
• This approach enables imaging speeds up to 32 times faster than current methods.
• The physics-informed design ensures accurate spectral separation and improved clinical applicability.