Convolutional Autoencoder for Automated Pre-Processing of Tumor Cell and Tissue Raman Spectra

Area of Science:

• Biomedical Optics
• Computational Biology
• Spectroscopy

Background:

• Raman spectroscopy (RS) offers label-free molecular profiling for analyzing tumor response to radiotherapy.
• Effective spectral pre-processing is crucial for accurate RS data analysis, involving baseline subtraction, smoothing, and artifact correction.
• Current pre-processing methods can be time-consuming and require manual intervention.

Purpose of the Study:

• To develop a single-step, automated spectral pre-processing method for Raman spectroscopy data from tumor cells and tissues.
• To evaluate the performance of a convolutional autoencoder (AE) in removing spectral artifacts and improving data quality for radiotherapy studies.
• To assess the AE's utility in identifying poor-quality spectra for semi-automated outlier detection.

Main Methods:

• A convolutional autoencoder (AE) architecture was employed for automated spectral pre-processing.
• Two AE models were trained: one for preclinical (cell line, xenograft) and one for clinical (prostate biopsy) Raman spectra.
• The AE's performance was quantified using root mean squared error (RMSE) and percentage root mean squared difference (PRD) compared to a baseline algorithm.
• A reconstruction AE was trained for semi-automated identification of poor-quality spectra.

Main Results:

• The AE demonstrated rapid and effective removal of baseline, noise, and cosmic rays (CRs) from both preclinical and clinical spectra.
• For preclinical data, AE achieved RMSE of 7.1 × 10^-5 and PRD of 3.1%, removing 94.0% of CRs.
• For clinical data, AE achieved RMSE of 8.1 × 10^-5 and PRD of 3.7%, removing 90.2% of CRs.
• The AE processed ~11,000 spectra in 2.4 seconds without a GPU.
• A reconstruction AE achieved 96.4% agreement with manual outlier identification.

Conclusions:

• The developed deep learning framework provides an efficient and automated solution for pre-processing tumor Raman spectra in radiation response studies.
• The AE significantly enhances data quality, enabling consistent extraction of biochemical radiation response profiles.
• This automated approach facilitates large-scale analysis of Raman spectroscopy data for clinical applications.