Glioma margin detection based on adaptive endmember perturbation and sparse unmixing with fiber Raman spectroscopy

Detection of glioma margin can assist surgeons in accurately determining surgical plans and improving surgical effects. Glioma growth is a process of infiltration from the tumor center to surrounding normal tissues. Thus, marginal tissues are mostly mixtures of malignant and normal tissues. In existing research, pattern recognition methods are widely used. However, their accuracy in margin detection is limited because they are optimized for classifying pure substances rather than mixed tissues. Therefore, a spectral unmixing method was employed for glioma margin detection. However, human tissue heterogeneity causes endmember perturbation, leading to low detection accuracy. A novel abundance estimation algorithm, Adaptive Endmember Perturbation Sparse Unmixing (AEPSU), is proposed in this study. The algorithm constructs an over-complete endmember library through clustering, employs endmember perturbation sparse unmixing algorithm to calculate the abundance matrix, and designs a progressive multiple spectral window unmixing integration method. By introducing the concept of Nash equilibrium, a weight calculation formula based on the standard replicator dynamic equation is proposed to update the weights of abundance vectors obtained from different spectral windows. The final abundance matrix is obtained via weighted fusion, thereby minimizing reconstruction error and enhancing unmixing accuracy. The method was evaluated using 112 positive spectra from glioma-margin tissues and 60 spectra from white-matter tissues. Comparative experiments with existing algorithms show that the proposed approach achieves an overall accuracy of 94.77%, outperforming the pattern recognition algorithms and unmixing methods adopted in the experiments. This paper demonstrates that the fiber-optic Raman spectroscopy has potential for intraoperative, noninvasive and rapid margin assessment.