Can artificial intelligence (AI) distinguish disease-related signals from patient-specific patterns in brain tumors? A recent study used deep learning (DL) to reduce batch effects in Raman spectroscopy data from glioma samples, retaining 81% balanced accuracy for IDH1 mutation classification while lowering patient-identification balanced accuracy from 68% to 12%. The findings, published in Scientific Reports, could support more reliable AI-assisted optical diagnostics.
Study: Deep learning mitigates batch effects in Raman spectroscopy. Image Credit: PUDRA/Shutterstock.com
Raman Spectroscopy Reveals Molecular Clues - and Hidden Bias
Raman spectroscopy is a non-destructive technique that uses laser light to identify the chemical components of biological tissue. As light interacts with molecular vibrations, some photons gain or lose energy, producing a spectrum that reflects the tissue’s composition, including proteins, lipids, and nucleic acids. Unlike methods that require chemical dyes, Raman spectroscopy can analyze tissue without adding labels or damaging the sample.
Combining Raman spectroscopy with AI has improved its use in medical diagnosis, including identifying brain tumors during surgery. However, the technique is sensitive to variations in laser power, room lighting, instrument settings, sample handling, and background fluorescence. These variations, known as batch effects, can cause AI models to learn differences in measurement conditions rather than focusing on disease-related signals.
This study therefore aimed to reduce the accuracy of patient identification, while retaining high accuracy in detecting mutations in glioma samples.
Deep Learning Targets Hidden Bias in Glioma Spectra
Researchers analyzed the APOLLO glioma cohort, which contained 300,506 Raman spectra from 59 tumor specimens across 46 patients. They
collected measurements using a Raman microscope with a 532 nm laser. The system utilized a 10× air objective and an electron-multiplied charge-coupled device detector to record spectra across 50–3400 cm-1.
To remove spectra affected by tissue damage and fluorescence, the study applied quality-control filters based on intensity quantiles, retaining 260,922 spectra for analysis. It compared three types of data: raw spectra, spectra denoised and baseline-corrected using the RADAR (Raman spectral analysis using DL for artifact removal) method, and manually baseline-corrected spectra. Each spectrum was normalized using min-max scaling.
To reduce patient-specific bias, a DL framework was developed combining a learnable feature-importance layer with one-dimensional convolutional neural networks (1D CNNs). The network architecture included batch normalization, leaky rectified linear units, max pooling, and output layers for patient identification and isocitrate dehydrogenase 1 (IDH1) mutation classification.
The model utilized training steps to distinguish disease-related signals from patient-specific spectral patterns. Initially, the feature-importance layer was held fixed while the convolutional network learned to predict both patient identity and IDH1 mutation status. Subsequently, the network weights were frozen, while the feature-importance layer was adjusted to minimize patient-identification performance while preserving the ability to classify IDH1 mutations.
AI Cuts Patient-Identification Accuracy From 68% to 12%
The outcomes indicated that raw Raman spectra contained strong patient-specific signals that could bias AI predictions. A model trained on uncorrected data distinguished between two glioma samples with identical mutation status, achieving an area under the receiver operating characteristic curve (ROC AUC) of 1.0 using both raw and manually baseline-corrected spectra.
Across the 46 patients, patient-identification balanced accuracy reached 68%. Conventional preprocessing methods had minimal impact: manual baseline subtraction reduced accuracy to 64%, while automated RADAR denoising lowered it to 66%.
The adversarial framework effectively reduced patient-identification balanced accuracy from 68% to 12%, while maintaining a balanced accuracy of 81% for IDH1 mutation classification, compared with 85% using raw spectra. It also narrowed the features used for classification to 392 spectral variables, primarily in the lower-wavenumber fingerprint region.
Download a free copy of this page here!
These features included a peak at approximately 532 cm-1 associated with cholesteryl esters, linked to changes in glioma metabolism. These findings suggest that the framework can effectively reduce patient-specific bias while preserving spectral information relevant to tumor classification.
Debiased Raman Spectra Retain 81% IDH1 Classification Accuracy
These findings could support the use of Raman spectroscopy in tumor assessment and point-of-care diagnostics. Probes used in neurosurgery can be affected by changes in lighting, probe movement, and laser alignment, which may introduce unwanted variations in spectral data.
By reducing these batch effects, adversarial debiasing could help AI models focus on disease-related signals. This approach may also facilitate the integration of Raman data collected using instruments, supporting the development of diagnostic models that work consistently across clinical sites.
Future Directions
This study suggests that combining adversarial machine learning with Raman spectroscopy could enhance the reliability of AI-based optical diagnostics by reducing patient-specific biases while preserving disease-related signals. This approach may support more consistent tumor classification across various clinical settings.
Future work should focus on testing the method across independent optical platforms, diverse cancer types, and different Raman techniques to assess its generalizability. Integrating debiasing methods into portable spectrometers could also support reliable, real-time analysis in clinical environments, including operating rooms.
Journal Reference
Sjöberg, J., et al. (2026). Deep learning mitigates batch effects in Raman spectroscopy. Scientific Reports, 16. DOI: 10.1038/s41598-026-62086-1, https://www.nature.com/articles/s41598-026-62086-1.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.