Raman spectroscopic and microscopic analysis of tissue type, molecular composition, and glioblastoma identification in brain tissue sections

Abstract

Glioblastoma (GB) is the most common primary malignant brain tumor. Despite improvements in treatments, survival probability has remained shorter than 2 years for most patients over the last 20 years. Accurate diagnosis of GB requires pathological evaluation of the tumor tissues using light microscopy, along with routine or specialized staining. Recent research also identified significant genetic/epigenetic alterations that influence diagnosis, prognosis, and treatment in addition to routine pathological evaluation. Identification requires the tissue to be sampled many times and analyzed using different methods that require additional time, resources, and expertise. To determine whether the tissue used for routine analysis can also be used to perform more detailed and comprehensive analysis without staining, we propose to use Raman Spectroscopy (RS), which is a label-free and non-destructive technique. RS provides molecule-specific spectra from the chemical composition of the sample for rapid analysis. In this thesis, we investigated GB, white matter (WM), gray matter (GM), and necrosis (NC) regions of GB patients using RS to determine whether a similar precision can be achieved as the routine histomorphologic diagnostic process. First, we proposed a refined protocol for effectively clearing paraffin from Formalin-Fixed Paraffin-Embedded brain tissue sections, without destroying the sample morphology and chemical composition, for eliminating the substantial Raman spectra of paraffin. We demonstrated that the less expensive and less toxic clearing agent CleareneTM removes paraffin as effectively as p-Xylene, the mostly used clearing agent in histopathology laboratories. Thus, we suggest substituting CleareneTM with p-Xylene for deparaffinization of brain tissue sections for Raman spectral analysis. Second, we optimized the choice of Raman spectrum acquisition parameters (excitation wavelength, acquisition time, accumulation count,), tissue thickness, and Raman substrate type (CaF2, glass). Third, we acquired the Raman spectra of GB, WM, GM, and NC regions and analyzed the spectral profile regarding the Raman peaks given in the literature. Raman spectra of GB and WM regions (nGB = 20, nWM = 18), which were annotated by an expert neuropathologist, have been classified with 87.2±1% GB and 90.7±1% WM training/test accuracies using machine learning models (SVM, kNN, RF). The effect of pre-processing of Raman spectra on classification accuracies has been investigated. Sample preparation conditions, Raman acquisition protocols, and machine learning classification models showed a successful proof-of-concept demonstration for the proposed Raman-based GB identification workflow. While there is room for further refining the machine learning models for improved training and validation accuracies, these protocols could be improved for eventual clinical utility. Once the clinical applicability and refined classification accuracies are demonstrated, these protocols might assist neuropathologists in error-free identification of GB in the clinics.