Key Points
- A review article in Photonics highlights how machine learning (ML) is being combined with optical spectroscopy techniques to improve the diagnosis and classification of central nervous system (CNS) tumors.
- These spectroscopic techniques enable real-time, non-invasive tumor analysis, but challenges such as complex, overlapping spectral data are being addressed through spectral preprocessing methods that enhance ML model performance for accurate tissue differentiation.
- Future progress in spectroscopy-based diagnostics depends on building larger, more diverse spectral datasets and developing portable spectroscopic tools to enable real-time, intraoperative applications and support the emerging “optical biopsy” approach in neuro-oncology.
Recent advancements in spectroscopic instrumentation and data analysis have propelled the field of neuro-oncology forward, particularly when it comes to diagnosing central nervous systems (CNS) tumors. A recent review article published in Photonics explores these developments (1). Led by Tatiana Saveleiva, who is a researcher affiliated with the Prokhorov General Physics Institute of the Russian Academy of Sciences and the National Research Nuclear University at Moscow Engineering Physics Institute, the research team delves into how the application of machine learning (ML) techniques are being used to obtain the spectral data from CNS tumors both ex vivo and in vivo.
What is neuro-oncology?
Neuro-oncology is a branch of science that studies the brain and spinal cord neoplasms. Doctors in this space are highly specialized, concentrating on treating patients with brain tumors (2). Clinical diagnostics in this space involve using MRIs to determine the type of tumor, the stage, and what specific treatment makes sense for the patient (2).
How is artificial intelligence (AI) being used to improve diagnostics of CNS tumors?
Saveleiva and her team begin the review by highlighting numerous studies that document how ML algorithms and spectroscopy are being integrated in clinical diagnostics. A common thread from these studies is that using both ML and spectroscopy improves neuro-oncologist understanding of tumor characteristics, which leads to better outcome predictions (1).
Optical spectral methods have been applied more frequently to great results. Some of the main methods and techniques used include fluorescence spectroscopy, diffuse reflectance spectroscopy, Raman spectroscopy, and infrared (IR) spectroscopy (1).
What are the benefits and drawbacks of these techniques?
As highlighted in the review article, these spectroscopic methods offer several advantages and disadvantages when used for clinical diagnostics. Fluorescence and diffuse reflectance spectroscopy have seen widespread intraoperative use for guiding the removal of diffuse glial tumors (1). These techniques offer real-time visualization and assessment of tumor margins, thereby increasing surgical precision while preserving healthy brain tissue (1).
Molecular spectroscopy techniques, such as Raman and IR spectroscopy, are rapid and non-invasive techniques, which makes them useful in clinical applications. However, these techniques can also produce complex spectral data, which adds a wrinkle to their application in clinical diagnostics more broadly. They often result in overlapping signals as well (1). To overcome this challenge, numerous studies have highlighted the value of spectral preprocessing, which helps improve analyzing spectral data by reducing noise, correcting for instrument defects, and improving spectral features (3).
The authors emphasize that proper preprocessing, which includes baseline correction, normalization, and noise reduction, significantly enhances the performance of ML classifiers (1). When these steps are effectively implemented, ML algorithms can distinguish between tumor and healthy tissue, or even between tumor subtypes, with high accuracy (1).
What are the future directions and ongoing challenges of spectroscopy-based diagnostics?
Currently, the industry is facing several challenges with spectroscopy-based diagnostics. The main one is that there is a need for more large data sets that can train ML models (1). Right now, multimodal spectroscopy is being used to mitigate this issue. In multimodal spectroscopy, different types of spectral data are combined. This trend, in conjunction with the development of spectral mapping techniques, is helping to build richer data sets, which is resulting in more sophisticated diagnostic outputs (1).
Another future direction and ongoing challenge in clinical analysis is the ongoing development of more portable spectroscopic instrumentation. Currently, the miniaturization of spectroscopy equipment is making it more realistic to deploy these tools in real-time applications (1). However, more work needs to be done to improve these instruments. The researchers believe the ongoing development of portable instrumentation will play a role in advancing the “optical biopsy” approach, where non-invasive or minimally invasive spectroscopic evaluations provide clinicians with near-instantaneous insights into tumor pathology (1).
Saveleiva and her team show in their review article that ML and spectroscopy are being used in tandem to advance CNS tumor diagnosis and treatment, and the maturation of these technologies will only continue to propel clinical diagnostics forward.
References
- Savelieva, T.; Romanishkin, I.; Ospanov, A.; et al. Machine Learning and Artificial Intelligence Systems Based on the Optical Spectral Analysis in Neuro-Oncology. Photonics 2025, 12 (1), 37. DOI: 10.3390/photonics12010037
- Ivy Brain Tumor Center, What is Neuro-Oncology? Ivy Brain Tumor Center. Available at: https://www.ivybraintumorcenter.org/brain-tumor-care/brain-tumor-treatments/neuro-oncology/ (accessed 2025-06-30).
- Chu, X.; Huang, Y.; Yun, Y.-H.; Bian, X. Spectral Preprocessing Methods. In Chemometric Methods in Analytical Spectroscopy Technology; Bian, X., Ed.; Springer: Singapore, 2022; pp. 111–168. https://doi.org/10.1007/978-981-19-1625-0_4.
