GLASSR-Net: Glass Substrate Spectral Restoration Neural Network for Fourier Transform Infrared Microspectroscopy in the Fingerprint Region

Abstract

Fourier transform infrared (FTIR) microspectroscopy has emerged as a pivotal pathological tool, offering informative spectral biomarkers for numerous diseases. However, the dependency on specialized infrared (IR) substrates limits effective and widespread clinical translation. IR transparent bases like calcium/barium fluoride (CaF₂/BaF₂) are costly and fragile, while IR reflective bases cannot be used for microscopic screening due to their opacity to visible light. In comparison, 1 mm thick pathological glass substrates are cost-effective, reliable, and widely utilized in clinical pathology. Therefore, establishing a methodology for collecting high-quality FTIR spectra on glass substrates is highly desired and beneficial. Here, we develop a glass substrate spectral restoration neural network (GLASSR-Net) to restore the fingerprint absorbance spectra from glass-based spectra spanning the wavenumbers from 1800 to 1000 cm^–1. The model is trained and validated by acquiring input glass-based spectra and ground truth spectra, respectively, through FTIR raster scanning on contiguous tissue sections of papillary thyroid carcinoma (PTC) mounted on glass and CaF₂ substrates. The GLASSR-Net successfully restores the sample absorbance and accurately reconstructs the biochemical distribution in both the spatial and spectral domains. Furthermore, the biochemical signatures of PTC are effectively extracted and analyzed from the restored spectra with traditional spectral histology, indicating a decrease in amide I/II absorption and an accumulation of lipids and nucleic acids in cancerous regions. The proposed GLASSR-Net presents a novel framework for data collection, spectral restoration, and integration of traditional methodology in glass-based IR microspectroscopy, which facilitates the incorporation of FTIR microspectroscopy into clinical histological scenarios.