Computational Assessment of Spectral Heterogeneity within Fresh Glioblastoma Tissue Using Raman Spectroscopy and Machine Learning Algorithms.

Understanding and classifying inherent tumor heterogeneity is a multimodal approach, which can be undertaken at the genetic, biochemical, or morphological level, among others. Optical spectral methods such as Raman spectroscopy aim at rapid and non-destructive tissue analysis, where each spectrum generated reflects the individual molecular composition of an examined spot within a (heterogenous) tissue sample. Using a combination of supervised and unsupervised machine learning methods as well as a solid database of Raman spectra of native glioblastoma samples, we succeed not only in distinguishing explicit tumor areas-vital tumor tissue and necrotic tumor tissue can correctly be predicted with an accuracy of 76%-but also in determining and classifying different spectral entities within the histomorphologically distinct class of vital tumor tissue. Measurements of non-pathological, autoptic brain tissue hereby serve as a healthy control since their respective spectroscopic properties form an individual and reproducible cluster within the spectral heterogeneity of a vital tumor sample. The demonstrated decipherment of a spectral glioblastoma heterogeneity will be valuable, especially in the field of spectroscopically guided surgery to delineate tumor margins and to assist resection control.

Impact of Formalin- and Cryofixation on Raman Spectra of Human Tissues and Strategies for Tumor Bank Inclusion.

Reliable training of Raman spectra-based tumor classifiers relies on a substantial sample pool. This study explores the impact of cryofixation (CF) and formalin fixation (FF) on Raman spectra using samples from surgery sites and a tumor bank. A robotic Raman spectrometer scans samples prior to the neuropathological analysis. CF samples showed no significant spectral deviations, appearance, or disappearance of peaks, but an intensity reduction during freezing and subsequent recovery during the thawing process. In contrast, FF induces sustained spectral alterations depending on molecular composition, albeit with good signal-to-noise ratio preservation. These observations are also reflected in the varying dual-class classifier performance, initially trained on native, unfixed samples: The Matthews correlation coefficient is 81.0% for CF and 58.6% for FF meningioma and dura mater. Training on spectral differences between original FF and pure formalin spectra substantially improves FF samples' classifier performance (74.2%). CF is suitable for training global multiclass classifiers due to its consistent spectrum shape despite intensity reduction. FF introduces changes in peak relationships while preserving the signal-to-noise ratio, making it more suitable for dual-class classification, such as distinguishing between healthy and malignant tissues. Pure formalin spectrum subtraction represents a possible method for mathematical elimination of the FF influence. These findings enable retrospective analysis of processed samples, enhancing pathological work and expanding machine learning techniques.

Machine Learning-Assisted Classification of Paraffin-Embedded Brain Tumors with Raman Spectroscopy.

Raman spectroscopy (RS) has demonstrated its utility in neurooncological diagnostics, spanning from intraoperative tumor detection to the analysis of tissue samples peri- and postoperatively. In this study, we employed Raman spectroscopy (RS) to monitor alterations in the molecular vibrational characteristics of a broad range of formalin-fixed, paraffin-embedded (FFPE) intracranial neoplasms (including primary brain tumors and meningiomas, as well as brain metastases) and considered specific challenges when employing RS on FFPE tissue during the routine neuropathological workflow. We spectroscopically measured 82 intracranial neoplasms on CaF2 slides (in total, 679 individual measurements) and set up a machine learning framework to classify spectral characteristics by splitting our data into training cohorts and external validation cohorts. The effectiveness of our machine learning algorithms was assessed by using common performance metrics such as AUROC and AUPR values. With our trained random forest algorithms, we distinguished among various types of gliomas and identified the primary origin in cases of brain metastases. Moreover, we spectroscopically diagnosed tumor types by using biopsy fragments of pure necrotic tissue, a task unattainable through conventional light microscopy. In order to address misclassifications and enhance the assessment of our models, we sought out significant Raman bands suitable for tumor identification. Through the validation phase, we affirmed a considerable complexity within the spectroscopic data, potentially arising not only from the biological tissue subjected to a rigorous chemical procedure but also from residual components of the fixation and paraffin-embedding process. The present study demonstrates not only the potential applications but also the constraints of RS as a diagnostic tool in neuropathology, considering the challenges associated with conducting vibrational spectroscopic analysis on formalin-fixed, paraffin-embedded (FFPE) tissue.