Machine learning for predicting the glioma type of live cells using Raman spectroscopy
Zannat, Noor E (2025-12-10)
Machine learning for predicting the glioma type of live cells using Raman spectroscopy
(10.12.2025)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe20251215119305
https://urn.fi/URN:NBN:fi-fe20251215119305
Tiivistelmä
Background: Gliomas are heterogeneous brain tumors, including aggressive glioblastomas, where rapid and accurate diagnosis is critical. Standard histopathology is invasive and slow. Raman spectroscopy provides a non-destructive alternative, probing molecular composition with high resolution. Coupled with machine learning, it offers potential for fast, precise glioma characterization and personalized care.
Methods: We collected Raman spectra from 284 live cell culture samples across four classes: astrocytoma (Astro), oligodendroglioma (Oligo), glioblastoma (GBM), and non-tumor controls. Preprocessing included cosmic rays removal, and baseline correction with airPLS. Spectra were normalized (min–max), silent regions and low-quality signals were excluded. A two-staged pipeline was developed: (1) cell contour detection from spectral profiles and (2) mutation status classification. Several classifiers were tested, and permutation feature importance was incorporated to identify discriminative Raman frequencies.
Results: Preliminary results demonstrate moderate separation of glioma IDH1 mutation status, with XGBoost classifier achieving moderate predictive performance across the four cell classes. Key discriminative features were observed in spectral regions associated with protein, lipid, and nucleic acid vibrations, consistent with known metabolic and structural differences between glioma types.
Conclusions and Outlook: Our workflow combines Raman spectroscopy with advanced preprocessing and machine learning for glioma classification. Beyond moderate accuracy, the identification of biologically meaningful spectral features enhances interpretability. This approach could accelerate diagnostics, reduce invasiveness, and support personalized strategies in neuro-oncology. Future work will extend the framework by increasing the performance of the model, classifying glioma subtypes, patient-derived samples, and advancing minimally invasive glioma diagnostics.
Methods: We collected Raman spectra from 284 live cell culture samples across four classes: astrocytoma (Astro), oligodendroglioma (Oligo), glioblastoma (GBM), and non-tumor controls. Preprocessing included cosmic rays removal, and baseline correction with airPLS. Spectra were normalized (min–max), silent regions and low-quality signals were excluded. A two-staged pipeline was developed: (1) cell contour detection from spectral profiles and (2) mutation status classification. Several classifiers were tested, and permutation feature importance was incorporated to identify discriminative Raman frequencies.
Results: Preliminary results demonstrate moderate separation of glioma IDH1 mutation status, with XGBoost classifier achieving moderate predictive performance across the four cell classes. Key discriminative features were observed in spectral regions associated with protein, lipid, and nucleic acid vibrations, consistent with known metabolic and structural differences between glioma types.
Conclusions and Outlook: Our workflow combines Raman spectroscopy with advanced preprocessing and machine learning for glioma classification. Beyond moderate accuracy, the identification of biologically meaningful spectral features enhances interpretability. This approach could accelerate diagnostics, reduce invasiveness, and support personalized strategies in neuro-oncology. Future work will extend the framework by increasing the performance of the model, classifying glioma subtypes, patient-derived samples, and advancing minimally invasive glioma diagnostics.