1.18 Automated Classification of Glioblastoma Margins in Label-Free SRS Microscopy Images

S. B. Lewis2, M. Ji3, X. S. Xie3, D. A. Orringer1  3Harvard University,Department Of Chemistry And Chemical Biology,Cambridge, MA, USA 1University Of Michigan Health System,Neurosurgery,Ann Arbor, MI, USA 2University Of Michigan Medical School,Ann Arbor, MI, USA

Introduction: Glioblastoma (GBM) is the most common and most aggressive form of intrinsic intracranial malignancy. Its surgical treatment is complicated by its capacity to infiltrate the surrounding parenchyma, with biopsies of ostensibly-normal tumor margin revealing large numbers of tumor cells after H&E staining. Stimulated Raman spectroscopy (SRS) microscopy is a label-free technique which can deliniate fine structure in fresh tissue, including nuclei and axons, using relative abundances of lipid and protein as contrast. Here we have derrved a program which automatically quantifies the cellularity, axonal density, and overall lipid:protein ratio, and classifies the sample as normal brain, infiltrating tumor or dense tumor. 

Methods: Two normal and two mouse xenograft models of GBM were imaged with SRS microscopy, at 2930 and 2845 cm-1 (the CH3 and CHRaman peaks, respectively). 1682 regions of interest were collected from these four brains and processed in Matlab using the Image Processing Toolbox. Briefly, nuclei were detected in the CH3 channel using the H-maxima transform and classified according to their compactness and intensity. Axons (in the CH2 channel) were hilighted via convolution of the image with the Sobel edge kernel and solidified via image closure; thresholding of the result was conducted via Otsu's method and candidate blobs were sorted by eccentricity and Euler number. Discriminant analysis was then used to classify each region of interest as "normal," "marginal" or "dense core". 

Results: The model differentiated normal tissue from dense core with 100% sensitivity and 100% specificity. When challenged to distinguish between normal tissue, marginal samples and dense core, abnormal tissue was detected with and 95.02% sensitivity 100% specificity. 

Conclusion: This model holds promise for the rapid and automated differentiation of normal and tumorous tissue during resection of malignant brain neoplasms, even when such tissue is not distinguishable by eye. This distinction is apparent without labels, staining, or fixation. Furthermore, aspects of both this technology and this algorithm may be applicable to the microscale segmentation of many tumors whose margins are unclear, or whose complete resection is imperative.