Apollo: Raman-based pathology of malignant glioma

Document Type

Conference Proceeding

Publication Date


Publication Title

Neuro Oncol


BACKGROUND: DNA methylation is an essential component for integrative diagnosis in glioma. Methylation subtype prediction of gliomas is currently done via sample extraction of high-quality of reasonable amount of DNA (∼1ug), methylome profiling, followed by probe identification, curation and subsequent analysis via different random forest classifiers. However, the DNA methylation classification is not always available for all the samples. METHODS: Raman Spectroscopy performed of the regions of interest using 1mm2 FFPE tissue spots from 45 patient samples with LGm1 to LGm6 methylation subtypes. Spectral information was then used to train a convolutional neural network (CNN) and develop a prediction algorithm. 70 % of dataset - model training while the remaining 30% for validation. Supervised wrapper methods and random forests were used to identify the top 109 most discriminatory Raman frequencies out of 1738. RESULTS: We identified the most discriminatory features from these analyses and demonstrated that these frequencies show differential spectral intensities for these frequencies depending upon the glioma subtypes across the larger areas of the tissue. We compared the results of the Ward linkage clustering with the separation induced by the 'frequency criterion', an empirical observation that Raman spectra of tumor spots are characterized by intensities higher than 5000 on some of the frequencies from 1463 to 1473. For each of the 45 samples we ran Ward linkage clustering with a variable number of clusters (from 2 to 7), with the majority cluster corresponding to tumor spots and the others corresponding to (various types of) non-tumor spots. We found that the majority cluster matches very well the tumor spots characterized by the frequency criterion, The average accuracy over all samples was 90:3%, the average precision was 99:6% and the average recall was 90:2%. For most samples, two clusters were sufficient to distinguish between tumor and non-tumor spots with accuracy.

PubMed ID

Not assigned.





First Page