Machine learning methods for spectrally-resolved imaging analysis in neuro-oncology

To reduce the frequency of relapses after surgical removal a brain tumor, it is critically important to completely remove all affected areas of the brain without disrupting the functionality of vital organs. Therefore, intraoperative differential diagnostics of micro-areas of tumor tissue with their subsequent removal or destruction is an urgent task that determines the success of the operation as a whole. Optical spectroscopy has shown its advantages over the past decade when used as a tool for intraoperative metabolic navigation. And one of the most promising options for the development of this technology is spectrally-resolved imaging. Currently, methods of spectrally-resolved imaging in diffusely reflected light have been developed, for example, mapping the degree of hemoglobin oxygen saturation, as well as fluorescence visualization systems, for both endogenous fluorophores and special fluorescent markers. These systems allow rapid analysis of tissue by the composition of chromophores and fluorophores, which allows the neurosurgeon to differentiate tumor and normal tissues, as well as functionally significant areas, during surgery. No less mandatory are the methods of using spectrally resolved visualization based on mapping characteristics obtained from Raman spectra, but due to the smaller cross-section of the process, these methods are used ex vivo , as a rule, for urgent analysis of fresh tissue samples. In this paper, we focus on both the physical foundations of such methods and a very important aspect of their application – machine learning (ML) methods for image processing and tissues’ classification.